标签：#developertools

Ask why agents write bad notes and the answers are all about effort. The model was lazy. The prompt was weak. It did not think about what would be useful later. Write better notes and the problem goes away. This misdescribes the task. Writing a note is not summarizing what happened. It is placing a bet on a question that has not been asked yet, by a reader who does not exist yet, in a context nobody can see. The writer is asked to optimize for a target that is not merely unknown but unknowable at the time the decision is made . Write-time precision in agent memory is not a hard problem solved badly. It is an ill-posed problem being answered anyway, and the outputs look like bad estimates because there is nothing else they could look like. The fix is not a better writer. The fix is a different unit of storage. Part 1: The Asymmetry 01. Why Write Is Not Retrieve The asymmetry is structural rather than practical. Retrieval is evaluable. Given a query and a corpus, there is a fact of the matter about which items are relevant. You can label it, measure it, and improve against the measurement. This is why information retrieval has a century of methodology and agent memory keeps borrowing from it. Writing is not evaluable in the same way, because the value of a note is not a property of the note. It is a property of the pairing between the note and a future query. At write time, one half of that pair does not exist. Not "is hard to predict." Does not exist. There is no distribution to estimate because the events that will generate the queries have not happened. Insight: Write-time precision in agent memory is undefined, not hard. The writer is not making a bad estimate. It is being asked to optimize a target that has no value to estimate, because the future queries have not been sampled from a distribution that has not been fixed. A better guess does not exist, however carefully the writer tries. The information-science literature has had a name for this for sixty years: a

Six coding agents were running. I could not tell you what any of them had done. Not roughly. Not approximately. The output was there, the files had changed, and the honest answer to "which one did that" was a shrug. Three questions in particular had no answer: which run burned the tokens, whether they genuinely ran at the same time or merely started together, and whether two of them had quietly edited the same file. That last one is the expensive question. An agent working on the wrong file looks exactly like an agent working on the right one, right up until you read the diff. The thing that was already true Every one of those runners writes a transcript to disk while it works. Claude Code does. So do Cursor, Codex, Gemini CLI, Copilot CLI and Kiro. The record of what happened was sitting in my home directory the entire time, in six different formats, none of which I had ever looked at. So runlanes does not wrap anything. There is no SDK, no instrumentation step, no account, and nothing to start before the run starts. It reads what the runner already wrote. The consequence is the part I did not expect to matter as much as it does: it works on runs that already finished. Most tools in this space need you to have decided, in advance, that this particular run was worth watching. This one can answer a question you only thought to ask afterwards. npx runlanes That opens a console on 127.0.0.1:4180 for whatever project you are standing in. There is no configuration file to write first. What it actually shows Now is every live session, across every runner it found, with what the main conversation spent against what it handed to subagents. On the session that motivated the whole thing, that split was 8.3 million tokens of conversation against 2.1 million delegated, which was not the ratio I would have guessed. The parallelism figure is the one I keep coming back to. Peak concurrency was four agents. The share of elapsed time where anything genuinely overlapped was 9% . Four

In the fast-paced world of software development, every second counts. Seamless tool interaction is not just a convenience; it's the bedrock of high developer productivity . Even seemingly minor hitches, like a non-functional 'copy link' button, can subtly erode efficiency, leading to frustration and lost time. A recent GitHub Community discussion highlighted just such an issue, where a user reported that the 'Copy link' button in Pull Requests (PRs) was consistently failing, specifically when using the Arc browser on macOS. This isn't merely about a broken button; it's a window into the complex interplay between browsers, web APIs, and the essential tools we rely on daily. The Reported Problem: A Month-Long Frustration The original post by vovapyc detailed a persistent problem: the 'Copy link' button in GitHub PRs had been broken for at least a month. The user specified their setup: Arc browser, MacBook Pro M1 Pro, and macOS 26.2. For dev teams, product managers, and delivery leads, a recurring point of friction like this, preventing a quick share of a PR link, represents a tangible drag on workflow. Imagine the cumulative time lost across a team if every developer had to manually copy URLs from the address bar multiple times a day. GitHub's automated response, while a standard and necessary part of their feedback loop, acknowledged the feedback and assured the user that their input would be reviewed. However, it didn't immediately offer a solution or explanation for the bug, leaving the user, and potentially others experiencing similar issues, in limbo. Diagram illustrating the three gates: Secure Context, Document Focus, and User Permission, that must be passed for the Clipboard API to function.## The Expert Insight: It's Likely the Browser, Not GitHub The true insight, and the crux of this discussion, arrived from hoangperry . Their comprehensive breakdown suggested that the issue was almost certainly browser-specific rather than a core GitHub bug. This distincti

If you are an active user of AI-powered IDEs like Cursor, VS Code with Copilot, or Windsurf, you already know the sinking feeling of seeing this notification: "You have used 100% of your fast premium requests for this billing cycle." Suddenly, your snappy, context-aware coding assistant slows to a crawl or starts racking up expensive pay-as-you-go API bills. At the same time, you are likely paying $20/month for a ChatGPT Plus or Team subscription that sits underutilized in a browser tab. You use it for general questions, but it lacks direct, real-time access to your local codebase, forcing you to engage in a tedious dance of copying and pasting code blocks. What if you could bridge this gap? What if you could let ChatGPT Web do the heavy reasoning and planning using your local context, while saving your premium IDE tokens for fast auto-completions ? In this article, we’ll explore a highly novel, intermediate-level setup that does exactly this. By leveraging the Model Context Protocol (MCP) , Node.js , and secure Cloudflare Tunnels , you can route heavy code-planning tasks directly to your web-based ChatGPT Plus subscription safely and completely free of extra token charges. The Philosophy: Let ChatGPT Think, Let Your IDE Work When building complex software with AI, your workflow generally splits into two distinct phases: Reasoning & Planning (High Token Usage): This is where you ask the AI to read 10 source files, understand the architecture, design a new feature, or find a subtle bug. This consumes massive amounts of context window tokens. Execution & Autocomplete (Low Latency): This is where the AI writes single lines of code, refactors a function, or autocompletes your imports. This requires fast, inline API queries. Paying premium API rates (per token) for Phase 1 is incredibly expensive. This is where this open-source MCP bridge project shines. It exposes a read-only view of your local project as an MCP server. Your web-based ChatGPT (via custom GPTs or MCP int

Tired of Claude Code generating bizarre, overly dramatic jargon like "load-bearing spine"? You can fix this by enforcing Simplified Technical English (STE) in your system instructions or .claudemd files. This 1970s aerospace standard restricts vocabulary, forcing your AI agent to communicate in clear, direct, and highly actionable prose. "The load-bearing spine has hit a ceiling, and that is a significant foot gun with a large blast radius." If you have spent any time recently working with AI coding agents, you have probably stared at your terminal reading absolute gibberish like this, wondering: What on earth are you trying to tell me? I asked a straightforward technical question, and instead of a direct answer, I got a theatrical performance. It is incredibly tiring to translate AI metaphors back into plain English just to figure out which line of code actually broke. Fortunately, there is a remarkably elegant fix for this. The solution does not involve complex prompt engineering; instead, it leverages a fifty-year-old aerospace standard: Simplified Technical English (STE) . Why does Claude Code output weird technical jargon? AI models generate overly dramatic jargon because they are trained on vast internet corpuses where technical writing is often cluttered, metaphorical, and performative. To sound authoritative, the model indexes on complex vocabulary and metaphorical hand-waving instead of simple, direct statements. Imagine a scenario where your team is debugging a database lock. A human engineer would say, "The transaction is blocked." An AI model, eager to please and sound sophisticated, might describe it as a "temporal execution bottleneck causing systemic architectural paralysis." This happens because reinforcement learning from human feedback (RLHF) often rewards models for sounding smart and comprehensive. Without strict stylistic constraints, the agent defaults to verbose, metaphorical explanations that add cognitive load rather than solving your proble

skillcheck is a static analyzer for SKILL.md files, the format agents like Claude Code, Copilot, Codex, and Cursor use to load reusable skills. It validates frontmatter, scores description discoverability, checks file references, enforces token budgets, and flags cross-agent compatibility issues. No network calls, no LLM calls, no file mutations. Runs as a CLI, a GitHub Action, or a pre-commit hook. pip install skillcheck skillcheck skills/ Latest pass was hardening and accuracy, not features. Here's what changed and why. Description scores went up. Skills that were scoring low because the scorer was broken will now see a jump in scoring. Median across the reference corpus went from 75 to 90. --explain-score also now tells you which pattern hits or misses instead of just a number. The score exists to predict whether an agent will actually find and trigger your skill, so a scorer that under-credits good descriptions defeats the point. The fix was validated against real-world skills, and the separation held: filler still scores 28-65, well-written descriptions 85-100. Corrupt files now fail cleanly instead of crashing. Before, a bad history ledger or non-UTF-8 skillcheck.toml above the skill dumped a Python traceback. It's now a clear error naming the file and byte offset (exit code 2). Config discovery walks up the directory tree, so one bad file could break every scan under it. Now every untrusted read (ingest, history, config) goes through the same guard before parsing, so they all reject the same way. README has been corrected in regards to token estimates. Without tiktoken, expect roughly 20-30% over-estimation, so install the extra if you're near a budget limit. The offline heuristic feeds the budget checks and its accuracy had never actually been measured, just assumed. It's benchmarked against tiktoken across the full corpus now, and the documented numbers are the measured ones. pip install "skillcheck[tiktoken]" The rest of the pass is invisible on purpose: f

Photo by Microsoft Copilot on Unsplash TL;DR: In a poll of 138 developers, three‑quarters say Claude Code outperforms Codex for everyday AI‑driven coding, pointing to higher accuracy, deeper context awareness, and a smoother workflow. The AI‑coding battlefield has been dominated by OpenAI’s Codex for years, powering tools like GitHub Copilot and shaping how developers write code. Yet a fresh wave of feedback suggests a shift: Anthropic’s Claude Code is rapidly becoming the preferred assistant for many programmers. A recent survey of 138 software engineers—spanning startups, enterprise teams, and freelance coders—revealed that 75% now rely on Claude Code as their go‑to AI partner. What drives this migration, and what does it mean for the future of AI‑augmented development? Survey Overview and Key Findings The questionnaire targeted developers who regularly use AI code generators, asking them to rank their primary tool and rate specific workflow attributes. Respondents represented a broad skill spectrum, from junior developers to senior architects, and worked across languages such as Python, JavaScript, Java, and Go. Adoption rate: 104 out of 138 participants (75%) listed Claude Code as their primary AI assistant, while only 34 (25%) still favored Codex. Primary criteria: Accuracy of generated snippets, ability to retain long‑form context, and ease of integration into existing IDEs topped the list. Secondary factors: Cost efficiency, response latency, and the perceived safety of the model (fewer hallucinations) also swayed decisions. The data paints a clear picture: developers are no longer satisfied with a one‑size‑fits‑all approach. They want an AI that can understand the nuance of a multi‑file project, stay on‑topic across extended sessions, and deliver code that compiles on the first try. Why Claude Code Wins Over Codex Higher Accuracy and Fewer Hallucinations Respondents repeatedly highlighted Claude Code’s ability to generate syntactically correct, production‑re

Documentation OPEN SOURCE official documentation content, searchable and organized ...

A few months ago we shipped the epilot CLI , and it quietly became one of my favorite tools. One command, npx epilot , gives you every single epilot API operation in your terminal: entities, journeys, workflows, pricing, files, permissions, 50+ APIs. Interactive pickers if you're exploring, --json and --no-interactive if you're scripting. It also turned out to be a perfect match for AI agents like Claude. Agents are great at driving CLIs: they discover operations, read the help, make calls, parse the JSON. No custom integration or MCP server needed, the CLI is the integration. And because handing an agent live CRM access is a scary idea, the CLI ships with two safety nets, both enforced server-side: # A session that physically cannot write. The restriction is baked # into the token, so the bearer can't turn it off. epilot auth login --readonly # A token that additionally gets all PII anonymized in every response epilot access-token createAccessToken -d '{ "name": "AI agent token", "read_only": true, "anonymize": true }' Read-only plus anonymized means an agent can explore, analyze and report on your real org all day, and the worst it can do is read data it can't even de-anonymize. Now we've made the CLI even better. On top of the raw API commands, we added app facades : a set of high-level epilot app commands that take you from an empty folder to a working app installed in your org. And that's what this post is really about, because apps are where the fun is. What are epilot Apps? epilot is very configurable out of the box: journeys, workflows, automations, pricing. But at some point every team hits a wall, something the UI simply doesn't offer. A custom tab on the contact or opportunity page showing data from your own systems A whole custom page in the epilot navigation Your own block in the journey builder A widget in the end-customer portal A flow action that calls your API when a workflow step runs An external product catalog or an API proxy to your backend That

If you've ever stared at a 15,000-file monorepo wondering "what actually breaks if I touch this...

There are two ways to work with an AI agent and I had tried both. Write the thing yourself and hand over only the tedious parts. Or hand over the whole task and audit whatever comes back at the end. The first is slow. The second is fast right up until it is wrong, and by then the wrong thing is finished. I expected this series to be about forcing a third option into existence. Nine parts of making an agent follow a workflow it would rather skip. That is not what happened. I never had to enforce it once. The queue that started this had a payload contract nobody had verified, and each phase after that cost me something before it gave anything back. A plan that would not move until the risk register named the provider contract the brief had only guessed at. A build that missed nothing except what my own brief left out. A review that stopped handing back a feeling and started handing back a verdict on every requirement I had already called done. A matrix instead of a trusted green run. A rollback with a name on it before anything got called shipped. And a retrospective that would not let a lesson through until it had checked itself against the trail. Eight parts of that. What I did not expect was which part turned out to be automatic. The Fight I Expected Never Started By the time I finish writing a requirement, I already know roughly what it is going to cost. Most engineers do. You can feel the difference between a one-line fix and something that is going to touch four files and a migration before you have written a single line of it. What I assumed was that the agent could not feel that, and that policing the gap would be my job forever. Reminding it to run the chain. Catching it when it decided a spike was small enough to skip. It has not needed the reminder. Small bugs do not trigger a brief and a plan, and they should not. A standard requirement, a spike, anything long or cross-cutting, runs the full cycle in order. The classification lands where I would have put i

Bilibili's 412 Incident, Explained: How v1.5.0 Absorbed It In June 2026, Bilibili quietly began rejecting yt-dlp with HTTP 412 errors. Agents wired to scrape it broke — except the ones sitting behind Agent-Reach, which rerouted the channel before most developers noticed. Agent-Reach is a local, MIT-licensed capability layer that gives shell-capable coding agents live internet access by selecting and routing to upstream CLIs rather than proxying data itself . When Bilibili started 412-blocking yt-dlp in June 2026, v1.5.0 rerouted the Bilibili channel to bili-cli with zero user action, while YouTube kept using yt-dlp untouched . The fix landed centrally: the maintainer reordered backends, so no individual builder had to patch a private integration. Quick Answer: When Bilibili began returning HTTP 412 to yt-dlp in June 2026, Agent-Reach v1.5.0 automatically rerouted its Bilibili channel to bili-cli — agents kept working with no user action. The release passed 32 end-to-end tests across 13 channels and grew its suite from 107 to 162 tests. The framing shift matters: v1.5.0 describes itself as a capability layer, not a tool collection. Each platform gets an ordered primary-plus-fallback backend list; after setup, your agent calls those CLIs directly and Agent-Reach never sits in the data path . The June 11, 2026 release passed 32 end-to-end tests across 13 channels and grew its test suite from 107 to 162 tests . Platform Primary backend Fallback Web pages Jina Reader — YouTube yt-dlp — GitHub gh CLI — RSS feedparser — Bilibili bili-cli OpenCLI (subtitles) Twitter/X twitter-cli OpenCLI Reddit OpenCLI rdt-cli XiaoHongShu OpenCLI xhs-cli LinkedIn linkedin-mcp Jina Reader Global search Exa via mcporter — "capability layer: multi-backend routing + real doctor + OpenCLI" — Agent-Reach v1.5.0 release framing (source: Agent-Reach CLAUDE.md ). The behavior is easy to model. The following minimal snippet — which was executed and returns exit 0 — illustrates the "absorb and keep wo

Somewhere on your feed right now, someone is bragging about the app they built in a weekend, no engineering background, no team, just a prompt and a Saturday. The post always ends the same way. Look what I built without needing any of you. Anyone can build software now. That is the whole pitch, repeated in a hundred different captions this month alone. Here is what that post never shows you. The part where someone checks it. Not "does it run." Checks it. Someone who did not write it, looking for the version of it that fails, the input nobody thought to try, the assumption that was wrong in a way the builder was structurally the worst person to catch, because they were too close to their own idea to see the hole in it. That someone is not optional. It is the actual job. None of this is theoretical. A notification icon that, instead of opening a panel, closes the entire page and drops me back on an empty tab. A video call that disconnects mid sentence for no visible reason. A video that plays with the sound simply gone, until I restart it. I do not have a chart proving reliability across the industry is getting worse. What I have is a pattern I keep running into, on products built by some of the most resourced engineering organizations on earth. Why software engineering has more than one person in the room A developer writes the code. A reviewer reads it before it merges. QA tries to break it on purpose. A manager decides if it is actually ready, or just finished. None of these roles exist because engineers do not trust themselves. They exist because a single person, however good, cannot see their own blind spots. That is not a flaw in the person. It is a fact about how blind spots work. Ten sets of eyes exist so that the eleventh mistake gets caught before a million people hit it. We have already watched what happens when that layer disappears, and we did not need AI to run the experiment. We ran it with the spreadsheet. The spreadsheet already showed us the cost of

You correct someone once. Not perfectly, but they get it. Next time, they do not make the same mistake. That is not optimism. That is just how correction works, "with people". I worked with agents on that assumption for a long time before I even noticed I was doing it. The plan that never held Before I had a single written rule anywhere, I would open a new session and ask for a plan first. Resolve the edge cases before touching a line of code, I said. The agent would agree, in whatever way a chat window agrees, and go straight to implementation anyway. I corrected it. Same session, it adjusted. New session, next day, same repo, same everything except the chat history: straight to implementation again. Every single time! So I did what looked reasonable. I wrote the plan myself. I resolved the edge cases myself, the open questions, the gaps the agent skipped past on its way to code. ' Tedious ' is the polite word for it. I was doing the one task I brought the agent in to do, and calling it collaboration. The same recipe, again The second correction arrived the same way. Every repo had its own shape. A recipe, a standard, a way things were supposed to be built here and not there. I would explain it. Full session, good results, the agent following the standard like it understood the standard. New session. Same repo, sometimes the new repo. Explain it again. Word for word, close enough. It was not that the agent forgot how to code. It was that nothing from the last conversation traveled with it into this one. Nothing said in the chat survives it I kept treating this like a training problem. Say it clearer. Say it earlier. Say it with an example next time. None of that was wrong exactly. It was aimed at the wrong layer. The actual mistake was assuming correction compounds the way it does with a person. It does not. A person carries what you told them into the next conversation without being asked to. An agent starts the next session exactly where it started the first one.

Nobody has ever asked you to upstream your dotfiles. For thirty years that was the deal with every tool we touched: if you didn't like it, you changed it, and the change lived with you. Then the tools started writing the code, and the deal quietly ended. This essay is about why the deal is back on the table. Because the thing that used to make forking expensive — the labour — is exactly what agents just made cheap. In This Article The Workbench Instinct Then the Harness Era Arrived Forking Was Always the Escape Hatch Code Got Cheap What I Shipped Into My Fork A Feature Does Not Have to Be Useful Your Desire Is the Limit The Fork Is the Destination Now the Discipline Part Bring Back the Joy The Workbench Instinct Show me a developer who has never touched their config and I'll show you someone who hasn't started yet. Vim users brag about their init.lua the way woodworkers talk about a hand plane they've had for twenty years. Emacs people wrote a whole operating system inside a text editor because they could. VS Code won partly because it shipped an extension API and got out of the way. Dotfiles repos are public artifacts, starred and forked, because the setup is part of the craft. This isn't productivity theatre. Some of it is genuine need, some of it is fixing a specific annoyance that only you have, and a lot of it is just fun. All three are valid. The workbench is where the joy lives — and nobody ever waited for permission to alias a command. Then the Harness Era Arrived Then agentic coding tools showed up and quietly changed the shape of the deal. The best-in-class agent harnesses are increasingly vendor-controlled. Claude Code is a product, not a repo you can clone and rebuild. Google announced it's retiring Gemini CLI in favour of a closed-source successor. And note where the line falls: Codex CLI is Apache-2.0 and sitting right there on GitHub, but the Codex desktop app — the thing most people actually click on — is not. The terminal stayed open. The interface

You set up Puppeteer, navigate to a page, call page.screenshot() , and the bottom half of your image is blank placeholder boxes. Welcome to lazy loading. Most modern sites defer images and heavy content until the user scrolls. Your headless browser never scrolls. So those elements never load. Here's how to deal with it. The scroll trick The most common fix is to programmatically scroll down the page before taking the screenshot: async function scrollToBottom ( page ) { await page . evaluate ( async () => { const delay = ms => new Promise ( r => setTimeout ( r , ms )); const distance = 300 ; while ( window . scrollY + window . innerHeight < document . body . scrollHeight ) { window . scrollBy ( 0 , distance ); await delay ( 150 ); } window . scrollTo ( 0 , 0 ); }); } await page . goto ( " https://example.com " , { waitUntil : " networkidle2 " }); await scrollToBottom ( page ); await page . waitForTimeout ( 1000 ); await page . screenshot ({ fullPage : true }); The 150ms delay between scrolls gives IntersectionObserver -based lazy loaders time to trigger. Too fast and you'll scroll past elements before they start loading. That final waitForTimeout after scrolling back to top lets any remaining images finish rendering. Not elegant, but necessary. Why networkidle2 isn't enough You'd think waitUntil: "networkidle2" would handle this. It waits until there are no more than 2 network connections for 500ms. But lazy-loaded images haven't even been requested yet at that point — they're waiting for a scroll event that never happens. networkidle2 only helps with content that loads on page init. For scroll-triggered content, you need the scroll. The loading="eager" override Some sites use the native loading="lazy" attribute. You can override it before images load: await page . evaluateOnNewDocument (() => { Object . defineProperty ( HTMLImageElement . prototype , " loading " , { set : function ( val ) { this . setAttribute ( " loading " , " eager " ); }, get : function () { retu

The call-order change came back pass-with-risk. I read the recommendation, saw it had a name and a reason, and felt the task close. Then I looked at the row under it. How was this verified: not run. Nobody had run the queue. I had a label. I did not have proof. This is Part 6 of The Contract Think produced a brief. Plan produced a gate. Build executed inside it. Review scored every requirement against a verdict instead of an impression. Review reads the diff and the plan and decides whether one satisfies the other. It does not run the queue. It cannot. Its whole job is judgment about what the code should do. Test is where someone finally checks what the code actually does. I had been treating those two as the same step. They are not. Test asks one question, and a verdict is not the answer For every active requirement, Test asks how it was verified. Command run, manual QA, or a comparison against known-good output. One of those three, or a written reason none of them ran. Not a recommendation. Not a risk level. Evidence. I built the matrix against the plan's requirements and filled in each row. Most had a command behind them. The call-order requirement had nothing. The cell read not run, and it sat directly below a pass-with-risk that already carried a name and a reason. That name had almost been enough for me. A named risk feels handled. It is not. It is a risk with a label on it, waiting for someone to actually look. So I ran the queue Three notifications, all with a real reason to fire within the same tick. The scheduler picked them up and ordered them by priority instead of arrival. Two landed in the sequence the requirement wanted. The third jumped ahead of a lower-priority notification that was still mid-processing. The change worked almost every time. Under one timing condition, it did not. That is the gap a verdict cannot see. Review had marked the requirement partial because the wording left the mechanism open. Running it found a real failure inside the mech

Cursor is the first AI code editor I have used that feels less like an autocomplete plugin and more like a place to steer work. It does not write perfect software. It changes the rhythm: ask for a scoped change, review the diff, then tighten it by hand. This Cursor AI review is based on day-to-day developer tasks: reading unfamiliar code, editing React components, moving logic between files, writing tests, and asking the editor to explain errors from the terminal. The short version is simple: Cursor is excellent when a task crosses file boundaries. It is less convincing when you only need cheap inline completions. What Cursor Actually Is Cursor is a VS Code-based editor from Anysphere with AI built into the core experience. Extensions, settings, themes, terminal panes, source control, and the familiar layout are still there. The difference is that chat, agent-style edits, tab completion, codebase search, and model selection are treated as editor controls rather than add-ons. That matters in daily use. I found the chat panel most useful when I pointed it at a directory and asked for a narrow change, such as "move this validation into the shared helper and update the tests." Cursor could usually find the right files, make a first pass, and leave me with a readable diff. I still had to check naming, edge cases, and test coverage, but it saved the boring part of hunting through files. The Best Part: Multi-File Editing Cursor's strongest feature is multi-file editing with codebase context. A lot of AI coding assistants can finish a function. Fewer can update the component, the hook, the type definition, and the test in one pass without losing the shape of the project. In my experience, Cursor is at its best with medium-sized tasks. It handles "add a field to this form and wire it through the API call" better than "invent a new architecture." It also works well for cleanup: renaming a concept, extracting repeated logic, or adding a missing test around an existing pattern.

GitHub's Agentic Workflows preview has the kind of headline that makes people reach for the wrong conclusion. Natural language Markdown can turn into GitHub Actions workflows. That sounds like "the YAML is going away." I do not think that is the interesting story. The interesting story is that the agent is not escaping the workflow engine. It is being pulled into it. That matters because a lot of agent demos still pretend the future is a smart process floating above the boring machinery: the agent understands the request, edits the repo, runs some commands, and hands back a neat result. Nice demo. Very clean. Production engineering is not clean like that. Production engineering has permissions, logs, runner groups, approval rules, secrets, firewalls, budgets, weird old repositories, compliance questions, and someone who has to explain what happened when the helpful automation did something surprising. So the shape of Agentic Workflows is useful precisely because it is less magical than the demo version. GitHub is putting agents inside the same CI/CD world that already carries a lot of organizational trust. That is the right direction. markdown is not the control plane The cute part is that a developer can describe a workflow in Markdown and have GitHub turn that into standard Actions YAML. That is useful. YAML is not a personality test, and most teams have better things to do than memorize every Actions syntax edge case. But Markdown is only the input surface. The control plane is still Actions. That distinction matters. If the generated workflow is a normal Actions workflow, then all the existing machinery can still matter: repository permissions, runner selection, logs, environments, approvals, branch protection, organization policy, and whatever security controls the company already built around CI. This is where I get more optimistic about agentic tooling. The bad version of agents asks every organization to trust a new, parallel execution model because the mode

Coding agents are no longer just autocomplete with a longer prompt. GitHub describes Copilot cloud agent as software that can research a repository, create an implementation plan, make code changes on a branch, run in an ephemeral GitHub Actions-powered environment, and let a developer review or create a pull request afterward. OpenAI's Codex GitHub integration similarly positions code review as a repository-aware review pass that follows AGENTS.md guidance and focuses comments on serious issues. That shift changes the buyer question. The useful question is not "does the agent usually write code?" It is "can the team detect when the agent drifts away from the developer's intent before the change reaches production?" A May 2026 arXiv paper, "How Coding Agents Fail Their Users" , gives teams a better vocabulary for that review. The authors studied 20,574 real IDE and CLI coding-agent sessions across 1,639 repositories and define misalignment as a breakdown that becomes visible through developer correction or pushback. The paper reports seven recurring symptom categories: wrong project diagnosis, misread developer intent, developer constraint violation, self-initiated overreach, faulty implementation, operational execution error, and inaccurate self-reporting. Effloow Lab also ran a bounded OpenAI API check using three synthetic, non-confidential coding-agent transcript snippets. The run did not measure real-world incidence, compare vendors, or reproduce the paper. It produced a small rubric that maps visible symptoms to review gates such as diff-scope checks, evidence-before-edit checks, acceptance-criteria coverage, and verification-output requirements. The public lab note is available at /lab-runs/coding-agent-misalignment-failure-taxonomy-poc-2026 . This guide turns that research and lab output into a practical QA checklist for teams buying, piloting, or packaging coding-agent workflows. Why This Matters for Agent Buyers Coding-agent procurement often starts with p