Everything we've written about agents and context.
97 guides, answers and articles, plus 33 definitions, grouped by subject instead of by format.
Memory & knowledge
Keeping what a team and its agents know, across sessions and after people leave.
See all 20
Connecting company data to AI agents: how to choose
Pick by where the data lives, not by which protocol looks strongest. If it sits inside one product that ships a connector, use that. If it is spread across a wiki, a drive, and a CRM that agents read daily, one place has to unify and structure it first, or every agent re-solves the same problem.
AnswerKnowledge transfer when an engineer leaves: what to capture
Capture the decisions, not the documentation. What leaves with a senior engineer was never written down: it lived in tradeoffs, rejected approaches, and incident history. Documentation mandates fail because writing docs is unrewarded work that competes with shipping. What survives a departure is knowledge captured as a byproduct of work already happening, somewhere the next person's tools can read.
ArticleAI memory lock-in works until it blocks your own product
Anthropic merged Claude chat and Cowork memory on August 25, 2026. AI memory lock-in is a moat until it stands between a vendor and its own next product.
ArticleGhost memory: why agents can't tell what's true now
Ghost memory is when old, current, and changed facts sit in an agent's memory and get retrieved together. Deleting loses history. Labeling the edge works.
ArticleKimi K3's 1M context runs mostly on linear attention
Kimi K3's 1M-token window runs on Kimi Delta Attention: hybrid linear attention with a fixed-size state. Why cheap long context still needs context curation.
ArticleDemand paging for the AI context window
A 2026 systems paper found 21.8% of tokens in agent context windows are wasted. Demand paging treats the AI context window as L1 cache, not full memory.
Retrieval & search
Getting the right slice of a corpus in front of an agent: retrieval, chunking, graphs, ranking.
See all 18Is RAG still worth it, or should you fine-tune?
Keep the retrieval, and fine-tune only what retrieval cannot fix. Fine-tuning teaches behavior, not facts: models learn new facts from it slowly and hallucinate more as they do. Retrieval still wins on anything that changes, needs citations, or is scoped per user. Fine-tune when the model already knows enough and only the output shape is wrong.
AnswerIs a knowledge graph worth it, or is better RAG enough?
Usually a better retrieval pipeline first. Most pilots that miss links are failing at entity resolution and lexical matching, not at graph topology, and hybrid search with reranking fixes that for a fraction of the cost. Add a graph when your questions genuinely chain facts across systems, and expect roughly ten accuracy points, not a transformation.
AnswerShould you process context at upload or at query time?
Freshness is a refresh-schedule problem, not a placement problem. Process at upload anything the document alone determines, and leave query time for work that depends on the question. When the corpus changes daily, keep that split and reingest only what changed. Nightly full rebuilds are what actually leave the index behind.
ArticleAgentic retrieval techniques that hold up in production
Six agentic retrieval techniques backed by 2026 benchmarks: read enforcement, query decomposition, routing, split routers, verification, and step budgets.
ArticlePruning context made agents more accurate, not just cheaper
Pruning an agent's tool history raised task completion from 71% to 91.6%. What the summary rescued was not content but the agent's place in its own work.
ArticleWhat chunk size should you use for RAG?
Studies disagree on the best chunk size for RAG: one found 100 tokens with zero overlap won. What the research supports, and the bigger lever most guides miss.
MCP & tool access
Connecting agents to tools and data over the Model Context Protocol, and what that costs.
See all 21JSON mode changes what your model answers
Across 44 language models, asking for JSON collapsed answer diversity: distinct answers fell from 52 to 36 and the top answer rose from 41% to 64% of responses.
ArticleStateless MCP: the spec is dropping session state
The 2026 MCP release candidate goes stateless: no initialize handshake, no session ID, any server instance answers any request. What stateless MCP means.
ArticleMCP Tasks: long-running work as context offloading
MCP Tasks let a server return a durable handle instead of a blocking result, keeping a long-running tool call's interim state off the agent's context window.
ArticleYour MCP server is installed. Your agent isn't using it.
Connecting an MCP server is easy. Getting an agent to call its tools on the first relevant turn is where teams lose, and the cause is context.
ArticleWhat Anthropic's context engineering guides leave out
Anthropic's 2026 trilogy on context engineering, tools, and code execution with MCP each assume the same missing layer: the substrate where context lives.
Guide Draft blog posts from customer meetings
Build a blog-post drafter in Claude Code that turns Granola meetings into grounded drafts. Wire holds your voice and editorial rules. Granola holds the meetings.
Context cost & limits
Window size, token spend, compression, and what to do when the budget runs out.
See all 13MCP server context window cost: what to cut first
Disconnect servers before you optimize them. Tool definitions are paid on every turn whether the agent calls them or not, so removing servers nobody uses is usually the largest single saving. After that, gate the remaining tools per session, defer schema loading until a tool is selected, and move occasional capabilities out of the tool list entirely.
ArticleContext window blindness: agents can't see their limits
Context window blindness: four frontier models misjudged their own context size by 43 to 84%. Why compaction drops the wrong things, and what fixes it.
ArticleContext engineering: what replaces prompt engineering
Prompt engineering has a new successor: context engineering. Learn why Karpathy and Tobi Lütke made the switch, and what it means for production AI systems.
ArticleContext pruning helps agents, until it doesn't
Context pruning helps AI agents in one regime and hurts in another. A 2026 study of models from 4B to 284B maps when to prune stale context and when not to.
ArticleContext bloat: why long-running agents break
Context bloat is when accumulated tool-call output crowds out an agent's task. Tool calls, not window size, break long-running agents. Here is the fix.
ArticleHow agents manage their own context window
Agent context management is shifting from fixed harness rules to learned, runtime decisions an agent makes about its own window. What the 2026 research shows.
Coding agents
Agents working in a codebase: what they can see, what they forget between sessions, and why.
See all 5AGENTS.md: is it enough for codebase context?
Re-explaining is a retrieval failure, not a memory failure. Every session starts cold, and the file you wrote your conventions into is either loaded whole, crowding out the working window, or never read at all. What ends it is recording the decisions your code cannot show, somewhere any agent can query on demand.
ArticleFile-native agents: when reading files backfires
A 9,649-experiment study found file-native retrieval lifts frontier-model accuracy 2.7% and drops open-source accuracy 7.7%. Match architecture to the model.
ArticleConstraint decay: structural rules break AI coding agents
Constraint decay: AI coding agents lose 30 points of accuracy under architecture and database rules. New EURECOM study explains why and where it hurts most.
ArticleWhat 466 AGENTS.md files teach about context engineering
An MSR 2026 study of 466 open source projects maps the five modes developers use to write AGENTS.md context, and what 50% file staleness reveals about practice.
Security & permissions
Scoping what agents can reach, and the attack surface that arrives with tool access.
See all 6Definitions
AI agent permissions: how to scope access down
Scope access per task, not per agent. Give each agent a narrow, short-lived credential for the job in front of it, keep irreversible writes behind an approval or a reversible path, and log what it actually reached rather than what it was allowed to reach. Standing broad grants are the failure mode, and 90 percent of deployed agents carry them.
ArticleMCP authorization decides what context agents see
MCP authorization became a context control plane in 2026. RFC 8707 token scoping decides which sources an agent can ever pull into its own context window.
ArticleWhy 14 MCP CVEs trace back to one design choice
OX Security's April 2026 advisory traces 14 MCP CVEs and 200,000 exposed servers to a single design choice: STDIO as the default local transport.
ArticleAI agents have too much access
88% of organizations report AI agent security incidents. The root cause is a context engineering failure: agents get all-or-nothing access, not scoped context.
ArticleTool poisoning: how MCP tool descriptions hijack agents
Tool poisoning hides instructions inside MCP tool descriptions the agent reads as trusted context. The MCPTox benchmark recorded a 72.8% attack success rate.
ArticleContext Poisoning: When Bad Data Becomes AI Ground Truth
Context poisoning plants false data into an AI agent's memory or RAG index. The model treats it as truth. It's a context engineering problem, not a model bug.
Multi-agent systems
Handoffs, sub-agents, and what breaks when several agents share the work.
See all 6Definitions
GPT-5.6 subagents: context isolation in the weights
GPT-5.6 Sol's Ultra mode gives each subagent its own context window. It solves context isolation but not handoff, and it multiplies your token bill fast.
ArticleAgent drift: why long-running AI agents lose the plot
Agent drift is how AI agents silently deviate from their goal over long tasks. Goal drift, stale context, and four more mechanisms cause it, not the model.
ArticleFive criteria of good context for AI agents
A 2026 paper formalizes five criteria for good AI agent context: relevance, sufficiency, isolation, economy, and provenance. Here's how to design for each.
ArticleSub-agent context isolation: the fix for context rot
Sub-agent context isolation gives each agent its own scoped window, stopping the context rot that kills multi-agent runs. Here's the pattern and its limits.
ArticleContext offloading: 3 patterns for AI agents
Context offloading keeps an AI agent's working context window small by moving state to a destination outside it. Three patterns, and what each one costs.
ArticleWhy every agent handoff corrupts your context
Every multi-agent handoff is a lossy compression event. Learn which five types of context degrade at agent handoff boundaries and how to preserve them.
Reliability & drift
Why agents degrade over long runs, and how that surfaces as confident wrong answers.
See all 4GPT-5.5 hallucination rate: what the numbers actually say
GPT-5.5's hallucination rate depends on grounding: 23% fewer wrong claims with tools, but 86% on AA-Omniscience without them. The real numbers, explained.
ArticleAI agent reliability is a context problem
AI agent reliability fails because the same task assembles different context every run. Non-determinism is a context engineering problem, not a model flaw.
ArticleAgentic context engineering: how ACE evolves contexts
ACE (ICLR 2026) beats tuned prompts by 10.6% with self-evolving contexts that avoid brevity bias and context collapse, two real failures of prompt tuning.
ArticleGPT-5.4-pro hallucinates more than GPT-5.4-nano
Vectara's 2026 benchmark shows OpenAI's flagship GPT-5.4-pro hallucinates at 8.3% while its nano variant stays at 3.1%. The reasoning-model tradeoff, explained.
Duplication is signal
Wire never skips or replaces a write. Every instance of the same content carries information, so we keep them all, and collapse only at retrieval.
EngineeringWhy we're moving Wire off Cloudflare Durable Objects
Four structural limits, not reliability: the vector index outside the object, compute far from the data, creation-time placement, and no self-hosting. The story behind Wire's new data plane.
ArticleMeta context engineering beats hand-tuned context
Meta context engineering (ICML 2026) learns the context-engineering process itself, beating ACE-style curation by 18 points while training 13.6x faster.
ArticleStructured Context vs Raw Text for AI
ETH Zurich found AI-generated context files hurt agent performance by 3%. Format choice alone swings LLM accuracy by 40%. Here's what the research says.
Browse by format
Give your agents one place to read and write.
Put in your docs, data and notes, then connect Claude, Codex or anything else with one link.
Create a container