标签：#agents

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

The front-running tax that bled crypto for a decade needs only observable intent and a party that controls order. Agent marketplaces are rebuilding both. In September 2020, a security researcher who goes by samczsun found about $12 million of someone else's cryptocurrency sitting in a vulnerable contract, exposed, and realized he had a few minutes to rescue it before someone less friendly noticed. He wrote the rescue transaction. Then he stopped, because he understood the problem with sending it. The moment his transaction hit Ethereum's public waiting area, the mempool, every bot watching that space would see a profitable move spelled out in plain code, copy it, pay a higher fee to jump ahead of him, and take the $12 million themselves. His rescue would become their heist, and he would have personally handed them the map. He wrote about this later in an essay called "Escaping the Dark Forest," borrowing a metaphor from Dan Robinson and Georgios Konstantopoulos at Paradigm, who had borrowed it from Liu Cixin's science fiction: an environment where any signal of your presence gets you killed, so the only survivors are the ones who stay silent and shoot first. The mempool is a dark forest. Broadcasting a valuable intention into it is detection, and detection is death. Samczsun survived only by refusing to play the open game. He submitted his rescue privately, straight to a miner, bypassing the public mempool entirely, so the predators never saw it coming. That story is usually told as a piece of crypto lore. I want to tell it as something else, because the thing that killed transactions in the dark forest was never really about blockchains. It was about a shape, and that shape is quietly being rebuilt inside the AI agent marketplaces that a lot of people are racing to launch right now. When it finishes, the same predators will be back, and this time the prey will be your agents. The three conditions, and why blockchain was just the extreme case The phenomenon samczsun

The one-year-old startup aims to reinvent financial architecture for autonomous AI transactions.

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

A team lead asks the question in a planning meeting and the room splits three ways: do we build this agent in Copilot Studio, write the orchestration ourselves and host it, or hand our container to Foundry and let it run our code? All three are official Microsoft build paths in 2026, all three end up in the same tenant-wide agent inventory, and the wrong pick costs you a rebuild once the project outgrows it. The answer is not "the most powerful one." It is the one whose service model matches who is building the agent, who owns the runtime, and how much pro-code control over orchestration and protocols you actually need. This article is the decision framework for that choice, grounded in Microsoft Learn and current as of mid-2026. Two of these three paths are public preview, so this is a guide to architectural fit and direction, not a production-reliability scorecard. TL;DR Three build paths, picked by service model, not power. Copilot Studio: low-code managed SaaS for makers. GA. Foundry Hosted agents: managed PaaS runtime for your own container. Public preview. Microsoft 365 Agents SDK: pro-code, self-hosted, widest channel reach. Agent Framework orchestrator in public preview. Monday move: before picking a platform, write down four things for this agent - who builds it (maker or pro-dev), who must own the compute, what channels it has to reach, and whether you need custom protocols or background/async behavior. Those four answers pick the path more reliably than a feature checklist. The three paths in one paragraph each Microsoft's own Cloud Adoption Framework frames the build options as three service tiers, which is the cleanest mental model to start from. The CAF positions them as Copilot Studio (SaaS, no/low-code), Microsoft Foundry (PaaS, pro-code or low-code), and GPUs and Containers (IaaS, code-first frameworks for maximum flexibility). The first two are managed by Microsoft. The third is where the self-hosted SDK path lives when you own the compute end to e

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

Your agent isn't hallucinating. It's reasoning correctly over the wrong inputs. Here's a failure pattern every team running agents in production has hit: An alert fires. The agent investigates: pulls metrics, checks recent deploys, scans logs, proposes a fix. The fix is confidently, articulately wrong. Instinct says blame the model (bad reasoning, needs a better prompt, maybe a bigger model). Then someone reconstructs what the agent actually saw. The metrics query returned a 5-minute-old cached aggregate. The deploy list was fetched before the relevant deploy landed. The log window was truncated at 1,000 lines and the line that mattered was #1,014. Given those inputs, the agent's conclusion was reasonable. It just wasn't debugging the incident that happened. It's garbage in, garbage out, with a twist that makes it worse for agents than for any previous software. A dashboard shows you the garbage. A human sees a stale chart and might notice the timestamp. An agent consumes the garbage silently and acts on it, with fluent reasoning layered on top. The output doesn't look like garbage; it looks like a confident, well-argued investigation. That confidence is what makes it dangerous. Why is this surfacing now For chatbots, context was mostly a retrieval problem over documents; mediocre retrieval meant a mediocre answer a human would shrug at. Three things changed with production agents: Agents act. A stale metric doesn't produce an off paragraph; it restarts the wrong service or pages the wrong team at 3AM. The cost went from cosmetic to operational. The input surface exploded. A production agent reads metrics, logs, deploy history, tickets, and chat from separate systems, each with its own latency, rate limits, caching, and clock. The "world" it reasons over is stitched together from partial snapshots, inside the model's reasoning loop, where nobody can inspect it. Errors compound. A single wrong input gives a slightly wrong answer. A 20-step investigation where step 3'

Perplexity added "Skills" to its Agent API, letting developers compose built-in and custom skills for more complex agent outputs. More skills mean more actions, and more actions mean more ways to produce an irreversible change. Before you add a fifth skill to your agent, make sure the first four have an undo path. The undo problem An agent skill that writes, sends, publishes, or deploys is an irreversible action if there is no rollback. The more skills an agent has, the more irreversible actions it can take in a single session. A chain of skills (research → draft → format → publish) can complete before a human notices the first step was wrong. The undo contract Every agent skill should declare: { "skill_name" : "publish_to_blog" , "action_type" : "write" , "reversible" : true , "undo_method" : "set_published_false" , "undo_timeout_seconds" : 3600 , "undo_side_effects" : [ "SEO index will retain the URL for up to 24h" ] } If reversible is false , the skill should require explicit human approval before execution. A skill registry with undo support # skill_registry.py """ Registers agent skills with undo metadata. Blocks irreversible skills from running without explicit approval. """ from dataclasses import dataclass from typing import Optional , Callable import json @dataclass class Skill : name : str action_type : str # "read", "write", "send", "deploy" reversible : bool undo_fn : Optional [ Callable ] = None undo_timeout_seconds : int = 3600 side_effects : str = "" requires_approval : bool = False class SkillRegistry : def __init__ ( self ): self . skills = {} self . execution_log = [] def register ( self , skill : Skill ): # Irreversible write/send/deploy skills require approval if not skill . reversible and skill . action_type in ( " write " , " send " , " deploy " ): skill . requires_approval = True self . skills [ skill . name ] = skill def execute ( self , skill_name : str , inputs : dict , approved : bool = False ): skill = self . skills . get ( skill_name ) i

Cursor's "Sand" project targets non-developers. Anthropic's Claude Cowork runs cross-device. OpenAI's ChatGPT Work delivers documents, spreadsheets, and presentations. When an agent produces a deliverable for someone who cannot read the code, the verification interface matters more than the output format. A download button is not verification. The problem When a developer reviews agent output, they can read the diff, run the tests, and check the types. When a non-developer receives an agent-produced spreadsheet or document, they have no equivalent verification path. They either trust it completely or reject it completely. Neither is useful. What a verification UI needs Element Purpose Implementation Source list Show what inputs the agent used Links to source documents with timestamps Confidence indicator Flag low-confidence outputs Per-section confidence derived from source coverage Change highlights Show what the agent generated vs. copied Diff view between source and output Audit trail Record who requested what and when Append-only log with request ID, timestamp, and result hash Rejection path Let the user say "this is wrong" without starting over Feedback record linked to specific output section A minimal verification component <!-- agent-output-verification.html --> <!DOCTYPE html> <html lang= "en" > <head> <meta charset= "UTF-8" > <title> Agent Output Verification </title> <style> .verification-card { border : 1px solid #ddd ; border-radius : 8px ; padding : 16px ; margin : 8px 0 ; font-family : system-ui , sans-serif ; } .source-list { list-style : none ; padding : 0 ; } .source-list li { padding : 4px 0 ; border-bottom : 1px solid #eee ; } .confidence-low { color : #c0392b ; font-weight : bold ; } .confidence-medium { color : #e67e22 ; } .confidence-high { color : #27ae60 ; } .reject-btn { background : #fff ; border : 1px solid #c0392b ; color : #c0392b ; padding : 4px 12px ; border-radius : 4px ; cursor : pointer ; } .reject-btn :hover { background : #c0392b

Visual Studio 2026's July update added organization-level custom agents for GitHub Copilot. An organization owner can now define a custom agent that every repository in the org automatically detects and offers in the agent selector. This is powerful and introduces a rollout risk: a single agent definition affects every developer in the organization simultaneously. The rollout problem When an org-level agent is published, every developer working in any repo in that org sees it in their agent selector. If the agent definition contains an incorrect instruction, a broken tool reference, or an overly permissive scope, it affects all developers at once. There is no canary by default. The agent definition review checklist Before publishing an org-level custom agent, review its definition against this checklist. 1. Instruction review # agent-definition.yml (example structure) name : org-code-reviewer description : Reviews PRs for security and style instructions : | Review each file change. Flag: - SQL injection in string concatenation - Missing input validation on public endpoints - Hardcoded credentials Do not modify files. Only comment. tools : - read_file - search_code - post_comment scope : organization Check: [ ] Instructions are specific enough to be testable ("flag SQL injection in string concatenation" not "review for security") [ ] Instructions do not conflict with existing repository-level AGENTS.md files [ ] The agent does not claim capabilities it does not have (e.g., "run tests" when no tools entry supports it) 2. Tool surface audit # Extract all tool references from the agent definition rg -n 'tools:' agent-definition.yml -A 20 | grep '^\s*-' For each tool: What resource does it access? (filesystem, network, repository API) Is it read-only or read-write? Does it require elevated permissions beyond the developer's own role? A post_comment tool can create noise across every PR. A write_file tool can modify code in every repo. Audit accordingly. 3. Canary test Be

Augment Code's Vinay Perneti talks models, harnesses, and context.

Thomas Betts talks with Clare Liguori, the technical lead on the open source Strands Agents SDK. The conversation covers how Strands Agents has grown from a Python SDK to a full agent harness running in production. Clare shares some lessons learned from building agents at scale, shifting to a model-driven architecture, and what comes next as the LLMs that underpin agents continue to improve. By Clare Liguori

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",