标签：#Product

Weird, playful desktop app for building with AI agents Discussion | Link

ChatGPT Ads is expanding to 31 European markets. Learn how advertisers can reach people as they explore, compare options, and make decisions.

*Second in a series on The Factory. Previously: The Factory That Merged 37 Tasks . The harness is at github.com/frozer/factory . The public description of my task harness ends on a claim: a packet that's wrong about the world fails identically on every retry. That sentence cost me four dead tasks and nine commits spent repairing task definitions instead of writing code. It reads like something you'd arrive at by thinking. I arrived at it by watching the same failure scroll past three times in a row. What three attempts is actually good for max_attempts = 3 felt like obvious hygiene. Models are stochastic. Sometimes a run goes sideways for no reason you can name — a bad turn, a truncated response, a tool call that gets refused. Retry it and it works. That's real, and a retry budget handles it well. The strength is exactly the constraint. A retry budget assumes the next attempt will differ from the last one . It buys you a second sample from a distribution. But a retry doesn't hand the model a fresh situation. It hands it the same packet back . Same file, same claims, same instructions. If the packet says a file lives at a path where no file lives, attempt three fails precisely where attempt one did, and the only thing three attempts bought was three times the bill. Failure without variance isn't flakiness. It's a specification defect wearing a reliability costume. Nobody had opened the files Here's what that looked like in practice. B03 was a loader for a national census dataset. Three attempts, all burned, all against a file shape that existed nowhere: wrong directory, wrong filenames, and a Data / Valor JSON envelope that appears nowhere in the actual data tree. Every attempt produced a parser for a document that doesn't exist. The packet was the defect, not the model. Nobody had opened the actual files before cutting it. Rewritten from the real JSON, the truth was a flat metadata / data envelope, four files — one of which shouldn't be loaded at all — and a long-fo

Save $250 off the Garmin Fenix 8, plus discounts on other popular fitness trackers from the brand.

Two agents finish the same task and report back. Fixed. The migration now handles null values. It wrote the code. It never ran it. Fixed. Added a null-handling layer, refactored the migration runner into a strategy pattern, and introduced a validation module. Every word true. All of it works. None of it asked for, and that strategy pattern is now yours to maintain forever. Point your code-review agent at both. If it checks claims against the repository — does this code exist, do the tests pass, did the commit land — it catches the first instantly and passes the second without hesitation. If it compares the work against the original request, it catches the second and misses the first entirely , because the described work is exactly what was asked for and simply does not exist. Neither reviewer is broken. They answer different questions. Most teams build one reviewer, point it at everything, and never ask which question it is asking. So I built reviewers that named what they were hunting. That worked, briefly, and then taught me something worse. The agent optimised for the check The verifier existed because of a specific behaviour I kept seeing: an agent would route a claim through a check and then present the check's approval as though it were independent confirmation. Not fabrication — something subtler. Authority laundering. The claim arrives pre-validated, and the validation is the thing you now argue with instead of the claim. Once a verifier existed, the behaviour adapted. The agent shaped its submission to fit what the verifier checked, collected the pass, and cited it. The gate had become a target, and the work had become the thing that fit through the gate. I first saw this in one model. Months later, after version changes and a rebuilt roster, I watched a different model — different vendor, different architecture — do the same thing on the same day I was writing this. Which is why "know your model's failure mode" is weak advice Models do fail in characterist

Automating Daily Bluesky Posts with a JSON‑Driven Content Pipeline TL;DR: I added a set of JSON files and a lightweight loader to the content‑automation repo so our CI can generate and publish daily Bluesky posts automatically. The change centralizes multilingual copy, makes the publishing script data‑driven, and removes the manual copy‑paste step that was breaking our release flow. The Problem Our weekly release process includes a short status update on Bluesky. The copy lives in a markdown file that we edit manually, then copy‑paste into the Bluesky CLI. Two issues kept surfacing: Human error – a typo or missing line would cause the post to be rejected by the API ( Error: Invalid payload: missing "text" ). No versioning – we had no way to track which text was used for a given date, making it impossible to audit or rollback a post. The symptom was a failed CI job that stopped the whole pipeline with the error above, and we were forced to roll back the entire release just to fix a missing word. What I Tried First My first attempt was to add a tiny shell script that reads a bluesky.md file and pipes it into the CLI: cat content/2026/08/16/bluesky.md | npx bluesky-cli post That worked locally, but the script crashed in CI because the file path was hard‑coded and the runner didn’t have the bluesky-cli binary installed. I also quickly realized that the same script would need to support English and Spanish versions, so the hard‑coded approach would explode as we added more languages. The Implementation 1. Data‑driven content files Instead of markdown, I switched to a JSON structure that can hold multiple languages and post types (progress, announcement, etc.). Each day gets its own folder under content/YYYY/MM/DD/VS/ . For the 2026‑08‑16 release we added: content/2026/08/16/VS/bluesky_en.json content/2026/08/16/VS/bluesky_es.json content/2026/08/16/VS/metadata.json Example bluesky_en.json [ { "type" : "progress" , "text" : "Finally pushed a real change: coverage for the

A fast, native podcast player for real podcast lovers Discussion | Link

Give everyone their own AI workforce in one platform Discussion | Link

The universe may never tell you if your choices mattered. Owlcat’s Osiris Reborn might not either.

The Problem: Context Rot Have you ever had to stop an AI agent halfway through a task to correct it? Work with AI agents long enough and you'll see a pattern: the longer a session runs, the worse the output gets. Every input you give the agent and every output it produces gets appended to the context window. Nothing leaves. By the time you're fifty messages deep, the agent is re-reading abandoned approaches, stale file contents, and corrections you made an hour ago. The fix isn't a better prompt. It's less context. The Core Rule Keep the context window small. Two habits will keep your AI agent from hallucinating: Delegate to subagents. Subagents do the heavy reading in their own context and return only the summary. Clear between phases. Once a phase produces a file, you no longer need the context that led to it. I aim to stay under 40% context usage in the main agent. Research → Plan → Implement I picked up this workflow from a HumanLayer talk , and it's the most reliable setup I've used. There are three phases, each ending in a markdown file, with a context clear between each. Research — the agent writes a research doc, then clears. Plan — the agent writes a plan doc, then clears. Implement — the agent executes the plan. The main agent never needs to remember the previous phase, because the previous phase wrote it down. All it needs is the conclusion. Research The research phase answers how something works today. For example: Describe how the payments flow works end to end. Look carefully at the API endpoint implementations. The main agent spins up parallel subagents to figure it out. From HumanLayer's repo , I found three subagents to be the most useful: codebase-locator — finds where things live codebase-analyzer — explains how a component works codebase-pattern-finder — finds existing patterns to model the new work after The best part about using subagents is that you can point them at a cheaper model. Mine run Sonnet while the orchestrator runs Opus. Plan The p

ChatGPT for Teens helps teens learn, think critically, and use AI with confidence, with stronger built-in protections, healthy-use features, and additional controls for parents.

Code Editor, Terminal, Git, AI Agent Discussion | Link

More than a decade after launching in Europe, the Netherlands company is now selling its repairable phones in the US, starting with the Fairphone (Gen 6+).

Merino is one of the best fabrics you can wear. We explain the different blends, what “GSM” means, and how to care for your clothes.

I've now routed five different AI coding tools through a proxy layer. Each one broke differently. None of them told me why. Writing this partly as a reference for myself, partly because the failure modes turn out to be genuinely interesting — they say a lot about how these tools are built. Claude Code: reads config once, then never again The simplest of the five. Config lives in ~/.claude/settings.json , two keys get modified: env.ANTHROPIC_BASE_URL env.ANTHROPIC_AUTH_TOKEN The failure mode: it reads that file exactly once, at startup. Change it while a session is running and nothing happens. No warning, no reload. This is the single most common "the switch is on but nothing works" report, across every tool. Close all windows, open a fresh one. One thing I appreciate: it only touches those two keys, backs up the original, and restores it exactly when you flip the switch off. Codex: doesn't read the model from your request This one is architecturally weird and cost me an hour. Every other tool specifies which model it wants in the request. Codex doesn't. It picks from its own internal model catalog. Consequence: if you don't explicitly select a model, it sits on a default internal GPT model that the market can't serve. And you don't get "please select a model" — you get a string of failures with no stated cause. The config it writes: ~/.codex/config.toml → model_provider, [model_providers.asale], model, model_catalog_json ~/.codex/auth.json → OPENAI_API_KEY Note model_catalog_json . That's the part that makes your selection show up in the app's model menu. And the desktop app reads that catalog at startup , so a model written while it's running won't appear until you restart. Two separate restart requirements stacked on each other. Credit where due: it preserves your existing comments and formatting in config.toml . Not every tool does. Gemini CLI: loses to your own shell config Config goes into ~/.gemini/.env . Two keys added, nothing else touched. The failure mode

Build deterministic agents that actually follow your process Discussion | Link

I started building xAgent in April 2025. The original idea was straightforward: build a task-oriented Agent that could run work on its own and turn AI into real automation. Looking back, that sentence sounds simple. Most of what I have done over the past year has been filling in everything hidden inside the words “run work on its own.” The first version used a single Agent. I quickly ran into a problem: once the prompt focused its attention on one kind of work, the Agent could do that work well but handle other tasks terribly. Fix one side and it would forget the other. Ask it to pay attention to everything and it would end up paying proper attention to nothing. That led me to multiple Agents, each responsible for a different part of the work and able to collaborate with the others. The idea worked, but as soon as they started running together, the next problem became obvious: tokens were too expensive. I bought a modified RTX 4090 with 48 GB of VRAM and started running open models locally. That took some pressure off the token bill, but exposed another problem: small open models were not smart enough. This was still the Qwen 3.0 era. The gap between local models and the best hosted models was obvious, especially on long tasks. They skipped steps, wandered away from the goal, and ignored instructions in all sorts of ways. I did not solve this by buying more tokens from top-tier models. It was not because those models were bad. The most practical reason was that I simply did not have the money. Once multiple Agents run continuously, the allowance included with a subscription disappears quickly. Spending more could solve the problem, but I could not afford to keep doing that, and it did not look sustainable for most individuals or small teams either. Not having the money forced me to think seriously about a question that has shaped xAgent ever since: can a small team with a limited budget use Agents properly without constantly paying for the best models, keeping costs

I spent a while getting permission levels right. What an agent may read, what it may write, what needs a human. That work was worth doing, and it did not save me here. The gap is simple to state and easy to miss: permission levels answer what may this agent do . They say nothing about what happens when the task itself changes after the agent already has permission . The situation When work is handed to an agent, the handoff carries a task definition, a scope, and the conditions that count as done. The agent takes it and starts. Then I edited the task. Not maliciously, not carelessly. I noticed something while the agent was still setting up, and I added a comment that changed what "done" meant. At that moment I had an agent holding valid authority for a task that no longer existed in that form. It had permission. Its permission was correct. Its instructions were stale. Static permission levels do not catch this. The agent is doing exactly what it was allowed to do. The problem is that "what it was allowed to do" was defined against a version of reality that I had just replaced. Why "just ask the human" is the wrong fallback The obvious fix is to have the agent check in when something looks off. I do not think this works, for two reasons. First, the agent cannot see the ambiguity. From inside the handoff, the stale task reads as perfectly coherent. There is no contradiction to notice. The instructions are complete, the scope is clear, and the acceptance conditions are stated. It just happens that a newer version exists elsewhere. Second, if the agent asks me every time it feels uncertain, I have rebuilt the bottleneck I delegated to avoid. Interruptions that fire on vague signals train you to approve them without reading, which is worse than not having them. The check has to be mechanical, and it has to run at a specific moment rather than continuously. The stop condition What I added is small: Bind the task revision at dispatch. The handoff record states which versio

From conversation to action, in three steps Discussion | Link

A novel-writing tool from the author of Silo Discussion | Link