--- title: Agent Context Management description: "Agent context management is the discipline of deciding what an agent should know, when it should know it, how context should be stored, and how context changes should be tested. The corpus treats context as a scarce, high-leverage resource rather than a simple large-window problem." created: 2026-05-22 updated: 2026-07-08 timestamp: "2026-07-08T00:00:00Z" type: concept tags: [agents, context, memory, knowledge-management, architecture] sources: [raw/imports/knowledge/ai-agents/2026-03-16-openviking-context-database.md, raw/imports/knowledge/agentic-coding/2026-03-10-cicd-context-agentic-coding-tessl.md, raw/imports/knowledge/agentic-coding/2026-03-17-claude-code-architecture-handbook.md, raw/imports/knowledge/ai-agents/2026-03-03-hermes-agent.md, raw/articles/blogwatcher/2026-05-24/2026-05-23-latent-space-ainews-all-model-labs-are-now-agent-labs.md, raw/imports/knowledge/research/tessl/2026-05-06-evaluating-ai-real-world.md, raw/articles/blogwatcher/2026-05-24/2026-05-22-tessl-ai-native-devcon-ldn.md, raw/imports/knowledge/research/tessl/2026-05-20-coding-agent-failure-patterns.md, raw/imports/knowledge/research/tessl/2026-05-18-multi-agent-ai-fix-meko.md, raw/imports/knowledge/ai-agents/tessl/2026-05-11-three-kinds-ai-agent-browser.md, raw/articles/blogwatcher/2026-05-25/2026-05-06-tessl-webmcp-making-web-apps-faster-and-cheaper-for-ai-agents.md, raw/articles/blogwatcher/2026-05-25/2026-05-05-tessl-with-cosmos-augment-code-wants-shared-memory-and-context.md, raw/articles/blogwatcher/2026-05-25/2026-05-01-tessl-from-blind-spots-to-merged-prs.md, raw/imports/knowledge/ai-agents/latent-space/new-ai-infra-unicorns-exa-modal-turbopuffer.md, raw/imports/knowledge/ai-agents/latent-space/giving-agents-computers-ivan-burazin-daytona.md, raw/imports/knowledge/ai-agents/latent-space/2026-05-20-railway.md, raw/imports/knowledge/references/pragmatic-engineer/the-pulse-antigravity-20-takes-ide-out-of-its-new-ide.md, raw/imports/knowledge/ai-agents/simon-willison/datasette-agent.md, raw/imports/knowledge/ai-agents/simon-willison/datasette-agent-sprites-0.1a0.md, raw/imports/knowledge/ai-agents/simon-willison/datasette-agent-charts-0.1a2.md, raw/imports/knowledge/ai-agents/simon-willison/datasette-agent-0.1a3.md, raw/imports/knowledge/ai-agents/simon-willison/2026-05-20-google-io.md, raw/imports/knowledge/ai-agents/latent-space/2026-05-20-ainews-google-io-2026-gemini-35-flash-omni-spark-antigravity-20.md, raw/imports/knowledge/references/pragmatic-engineer/2026-05-19-ai-impact-on-software-engineers-part-2.md, raw/articles/blogwatcher/2026-05-28/2026-05-27-tessl-dont-make-your-agent-guess.md, raw/articles/blogwatcher/2026-05-28/2026-05-27-latent-space-ainews-new-ai-infra-decacorns-fireworks.md, raw/articles/blogwatcher/2026-05-28/2026-05-27-simon-willison-openai-anthropic-product-market-fit.md, raw/articles/blogwatcher/2026-05-28/2026-04-30-tessl-warp-goes-open-source.md, raw/articles/blogwatcher/2026-05-28/2026-04-29-tessl-openai-open-sources-symphony.md, raw/articles/blogwatcher/2026-06-15/2026-06-03-tessl-ai-native-devcon-day-2-from-agent-demos-to-operating-models.md, raw/articles/blogwatcher/2026-06-15/2026-06-14-simon-willison-why-ai-hasnt-replaced-software-engineers-and-wont.md, raw/articles/blogwatcher/2026-06-15/2026-06-03-simon-willison-uber-caps-usage-of-ai-tools-like-claude-code-to-manage-costs.md, raw/articles/blogwatcher/2026-06-15/2026-06-06-simon-willison-running-python-code-in-a-sandbox-with-micropython-and-wasm.md, raw/articles/blogwatcher/2026-06-15/2026-06-13-simon-willison-mapping-sqlite-result-columns-back-to-their-source-table-column.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-25-latent-space-ainews-its-meta-harness-summer.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-latent-space-why-the-frontier-ecosystem-must-be-open-matei-zaharia-and-reynold-xin-databricks.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-tessl-cursors-new-leaderboard-shows-teams-the-most-popular-plugins-skills-and-mcps.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-latent-space-ainews-claude-tag-multiplayer-proactive-persistent-agents-in-slack.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-23-tessl-why-warp-is-betting-engineering-leaders-are-done-picking-a-favourite-coding-agent.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-simon-willison-simonw-browser-compat-db.md, ../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-23-simon-willison-opfs-pyodide-test-harness.md] confidence: medium contested: false contradictions: [] --- # Agent Context Management ## Definition Agent context management is the discipline of deciding what an agent should know, when it should know it, how context should be stored, and how context changes should be tested. The corpus treats context as a scarce, high-leverage resource rather than a simple large-window problem. ## Recurring Patterns - Progressive disclosure: keep a small resident entry point and load details on demand. - Layering: resident instructions, path-loaded rules, on-demand [[agent-skills]], isolated subagents, and deterministic hooks. - Repository-local knowledge: context outside the repo or workspace is invisible unless deliberately imported. - Context CI/CD: changes to instructions, skills, and configuration need evals, observability, and change logs. - Memory systems: platforms such as [[hermes-agent]] and OpenViking preserve user, agent, and resource context over time. ## OpenViking Pattern The OpenViking source describes a file-system-like context database with tiered loading: L0 abstracts, L1 overviews, and L2 details. It is presented as an alternative to flat vector storage, with recursive directory retrieval and observable retrieval trajectories. ## Risks - Context can become stale without breaking tests. - Tool outputs can drown useful information in noise. - Adding instructions can shift model behavior globally. - Sensitive memory can create privacy and security risk when agents have broad account access. ## Related Pages - [[agentic-coding]] - [[agent-skills]] - [[claude-code]] - [[openclaw]] ## Blogwatcher Ingest Notes ### 2026-05-24 The Latent Space and Tessl items both point toward model-plus-harness systems: memory, workflow, UI, evaluation, and context are becoming part of the product surface. This updates the context-management page with another source cluster connecting [[agent-context-management]] to [[agentic-coding]], [[agent-skills]], and [[openclaw]]. ### 2026-05-26 The newest Blogwatcher sources add more evidence that context is moving into products and infrastructure: Daytona gives agents computer-like workspaces, Railway is framed as agent-native cloud, Google/Antigravity emphasize background agents, and Datasette Agent embeds an agent loop into a data application. This supports the existing pattern that [[agent-context-management]] spans UI affordances, persistent workspaces, tools, and runtime boundaries, not only prompt-window packing. ### 2026-05-28 New sources add concrete context-management pressure points: Tessl's tool-design article treats tool schemas and authority boundaries as context that shapes behavior; Latent Space/AINews highlights context governance, trustworthy memory, dynamic skill routing, managed harnesses, and sleep-like context consolidation; and Symphony/Warp/OpenAI-style orchestration pushes context into issue trackers, isolated workspaces, and review loops. Related: [[agentic-coding]], [[agent-skills]], [[agent-security]]. ### 2026-05-29 The Cognition/async-agent sources reinforce that context is now a product architecture problem: background agents need task specs, repo state, machines, memory, tool access, and review loops to persist outside the developer's foreground session. Simon Willison's Datasette Agent and LLM-limit notes add a smaller-scale implementation signal: agent products need explicit accounting for model limits, tool use, and application-specific state. Related: [[agentic-coding]], [[agent-evaluation]], [[agent-skills]]. ^[raw/articles/blogwatcher/2026-05-29/2026-05-28-latent-space-the-age-of-async-agents-cognition-s-walden-yan-openinspect-s-cole-murray.md] ^[raw/articles/blogwatcher/2026-05-29/2026-05-15-simon-willison-datasette-agent-0-1a2.md] ^[raw/articles/blogwatcher/2026-05-29/2026-05-15-simon-willison-datasette-llm-limits-0-1a0.md] ### 2026-05-30 Claude Opus 4.8's mid-conversation system messages and lower prompt-cache minimum make context steering itself a platform feature: long-running agents can receive updated instructions without restating the full system prompt, preserving cache behavior and reducing input cost. Dynamic Workflows and Datasette Agent also reinforce the trend toward context living in orchestrators, tools, and application state rather than only in a chat transcript. Related: [[agentic-coding]], [[agent-skills]], [[claude-code]]. ^[raw/articles/blogwatcher/2026-05-30/2026-05-29-tessl-ai-coding-agent-accuracy-opus-4-7-vs-4-8.md] ### 2026-06-13 This Blogwatcher batch adds a fresh cluster around coding-agent economics and governance: model billing/name mismatches, context-engineering cost, reviewability as the bottleneck, agent skills/harnesses, enterprise operating models, Claude Fable behavior, Claude Code misuse/security reporting, and Simon Willison's concrete Datasette/MicroPython/WASM sandboxing examples. Related: [[agent-context-management]], [[agent-evaluation]], [[agent-security]], [[anthropic]], and [[datasette]]. ^[raw/articles/blogwatcher/2026-06-13/2026-06-12-simon-willison-quoting-andrew-singleton.md] ^[raw/articles/blogwatcher/2026-06-13/2026-06-12-tessl-why-your-gemini-bill-doesn-t-match-the-model-names.md] ^[raw/articles/blogwatcher/2026-06-13/2026-06-12-tessl-claude-fable-5-vs-opus-4-8-the-mythos-hype-meets-reality.md] ^[raw/articles/blogwatcher/2026-06-13/2026-06-12-latent-space-ainews-loopcraft-the-art-of-stacking-loops.md] ^[raw/articles/blogwatcher/2026-06-13/2026-06-11-simon-willison-claude-fable-is-relentlessly-proactive.md] ### 2026-06-14 The selected sources add another context-management signal: hidden context costs, benchmark/reviewability pressure, and model/tool pricing all make context a production resource that needs design and evaluation rather than ad hoc prompt stuffing. Related: [[agentic-coding]], [[agent-evaluation]], and [[agent-security]]. ^[raw/articles/blogwatcher/2026-06-14/2026-06-13-latent-space-ainews-fable-and-mythos-officially-too-dangerous-to-release.md] ^[raw/articles/blogwatcher/2026-06-14/2026-06-13-simon-willison-statement-on-the-us-government-directive-to-suspend-access-to-fable-5.md] ^[raw/articles/blogwatcher/2026-06-14/2026-06-12-simon-willison-openai-webrtc-audio-session-now-with-document-context.md] ^[raw/articles/blogwatcher/2026-06-14/2026-06-10-latent-space-ainews-anthropic-claude-fable-5-mythos-but-safe-with-controversial-ter.md] ### 2026-06-15 Tessl's AI Native DevCon operating-model coverage and Simon Willison's engineering/provenance posts reinforce that useful agent systems depend on explicit operating context, inspectable provenance, and cost/review constraints rather than just stronger base models. Related: [[agentic-coding]], [[agent-evaluation]], and [[datasette]]. ^[raw/articles/blogwatcher/2026-06-15/2026-06-03-tessl-ai-native-devcon-day-2-from-agent-demos-to-operating-models.md] ^[raw/articles/blogwatcher/2026-06-15/2026-06-14-simon-willison-why-ai-hasnt-replaced-software-engineers-and-wont.md] ^[raw/articles/blogwatcher/2026-06-15/2026-06-03-simon-willison-uber-caps-usage-of-ai-tools-like-claude-code-to-manage-costs.md] ^[raw/articles/blogwatcher/2026-06-15/2026-06-06-simon-willison-running-python-code-in-a-sandbox-with-micropython-and-wasm.md] ### 2026-06-17 This Blogwatcher batch adds a fresh cluster around coding-agent economics and governance: model billing/name mismatches, context-engineering cost, reviewability as the bottleneck, agent skills/harnesses, enterprise operating models, Claude Fable behavior, Claude Code misuse/security reporting, and Simon Willison's concrete Datasette/MicroPython/WASM sandboxing examples. Related: [[agent-context-management]], [[agent-evaluation]], [[agent-security]], [[anthropic]], and [[datasette]]. ^[raw/articles/blogwatcher/2026-06-17/2026-06-16-simon-willison-datasette-tailscale-0-1a0.md] ^[raw/articles/blogwatcher/2026-06-17/2026-06-16-simon-willison-quoting-georgi-gerganov.md] ^[raw/articles/blogwatcher/2026-06-17/2026-06-16-interconnects-ai-frontier-post-training-recipe-review-with-finbarr-timbers.md] ^[raw/articles/blogwatcher/2026-06-17/2026-06-16-simon-willison-the-fable-5-export-controls-harm-us-cyber-defense.md] ^[raw/articles/blogwatcher/2026-06-17/2026-06-16-simon-willison-quoting-matteo-wong-the-atlantic.md] ### 2026-06-19 Tessl's Skill Inventory reframes agent context as a governed "skill estate": it scans a GitHub org, maps duplication/overlap/ownership drift, and surfaces fan-out risk from linked skills and evals. The "OpenClaw for Dummies" tutorial and Tessl's native OpenClaw skill support reinforce that instructions/skills are first-class, installable, versioned context. Charity Majors on disposable code and the "bright kid" evals-can-lie piece underline that context engineering (rules, skills, scripts) is now the binding constraint rather than raw model capability. Related: [[agent-skills]], [[agentic-coding]]. ^[raw/articles/blogwatcher/2026-06-19/2026-06-17-tessl-skill-inventory.md] ^[raw/articles/blogwatcher/2026-06-19/2026-03-27-tessl-openclaw-for-dummies.md] ^[raw/articles/blogwatcher/2026-06-19/2026-06-17-simon-willison-charity-majors.md] ^[raw/articles/blogwatcher/2026-06-19/2026-03-26-tessl-bright-kid-no-manners-part-2.md] ### 2026-06-20 Two sources sharpen the context/memory-as-estate thesis. (1) **Anthropic "auto dream"** (via Tessl) is a between-session maintenance pass for [[claude-code]]'s `MEMORY.md`: it reviews memory, removes stale/contradictory notes, merges overlapping entries, standardizes details (relative timestamps → fixed dates), and rebuilds the index — explicitly modelled on REM-sleep consolidation and operating only on memory files, never the codebase. It is distinct from in-session context compression and compaction; it cleans memory *between* sessions. (2) **Sean Lynch's HN comment** (via Simon Willison) reframes MCP's durable value as **isolating auth/credential flows outside the agent's context window — and potentially out of the harness entirely** ("maybe the idealized form of MCP is just an auth gateway"). Both reinforce that secrets, stale notes, and credentials belong outside model-reachable context. Related: [[agent-skills]], [[claude-code]], [[anthropic]], [[agent-security]]. ^[raw/articles/blogwatcher/2026-06-20/2026-03-26-tessl-anthropic-auto-dream.md] ^[raw/articles/blogwatcher/2026-06-20/2026-06-19-simon-willison-sean-lynch.md] ### 2026-06-26 Context management is moving from prompt files into **organizational control planes**. Omnigent and Warp/Oz treat context, session history, budgets and agent choice as portable harness-level concerns; Cursor’s Customize page/leaderboards expose which skills, MCPs and plugins are actually active in teams; Claude Tag raises the same issue inside Slack, where selected channels/tools/data/codebases become the agent’s operational context. Simon Willison’s browser-compat SQLite and OPFS/Pyodide experiments add a smaller local-first pattern: structured databases and browser-local files can become explicit context substrates for agents and Datasette Lite. Related: [[agentic-coding]], [[agent-skills]], [[datasette]]. ^[../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-latent-space-why-the-frontier-ecosystem-must-be-open-matei-zaharia-and-reynold-xin-databricks.md] ^[../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-tessl-cursors-new-leaderboard-shows-teams-the-most-popular-plugins-skills-and-mcps.md] ^[../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-latent-space-ainews-claude-tag-multiplayer-proactive-persistent-agents-in-slack.md] ^[../knowledge/raw/articles/blogwatcher/2026-06-26/2026-06-24-simon-willison-simonw-browser-compat-db.md] ### 2026-06-28 Tessl's World's Fair post makes context-management process explicit with a “Context Development Lifecycle”: recurring agent mistakes are mined from PRs, session logs, and tickets, converted into context/workflow changes, and shipped through the repo's normal automation path. This is a productized version of the page's existing claim that context needs CI/CD, observability, and change control, not ad hoc prompt accumulation. Related: [[agentic-coding]], [[agent-evaluation]], [[agent-skills]]. ^[../knowledge/raw/articles/blogwatcher/2026-06-28/2026-06-27-tessl-see-you-at-ai-engineering-worlds-fair-2026.md] ### 2026-07-02 Tessl's Microsoft/A2A summary adds a concrete multi-agent context-transfer pattern: keep domain agents stateless and send summarized conversation history in each message payload, with the coordinator deciding exactly what context each downstream agent sees. The tradeoff is explicit — summaries can lose or distort detail, but shared storage creates cross-team credentials/connectivity dependencies and stateful domain agents complicate deployment, rollback, and audit. This strengthens the page's claim that context architecture is also a governance and security-boundary decision, not just a prompt-window optimization. Related: [[agentic-coding]], [[agent-security]], [[agent-evaluation]]. ^[../knowledge/raw/articles/blogwatcher/2026-07-02/2026-06-29-tessl-context-travels-multi-agent-world.md] ### 2026-07-04 Vercel's Andrew Qu supplies a product-platform version of context management: eve bundles primitives for context, tools, resumability, sandboxes, long-running jobs, observability, and evals; skills correct stale model knowledge; and Vercel serves Markdown to detected agents as an agent-readable web surface. The AIEWF dispatch adds the organizational constraint: Anthropic's Claude Tag makes work more delegated, async, and proactive, but bottlenecks shift to review and humans' ability to conceptualize the system. Related: [[agentic-coding]], [[agent-skills]], [[claude-code]]. ^[../knowledge/raw/articles/blogwatcher/2026-07-04/2026-07-03-latent-space-vercel-agents-new-software.md] ^[../knowledge/raw/articles/blogwatcher/2026-07-04/2026-07-03-latent-space-aiewf-daily-dispatch-loops.md] ### 2026-07-08 Lilian Weng and BAIR both make context/state architecture a first-class substrate for agents rather than a prompt-window afterthought. Weng frames file systems, logs, subagent outputs, workflow state, skills, and context playbooks as harness components that survive beyond one chat; BAIR argues that MD-file memory and simple embedding retrieval will strain when thousands of agents coordinate, proposing structured, faceted corrective memory, shared state, rollback, and consensus/coordination mechanisms for agent swarms. Latent Space's Fable field guide adds the operator-facing version: blindspot passes, implementation notes, "interview me" prompts, and references as ways to surface unknowns and prevent the harness from hobbling a model. Related: [[agentic-coding]], [[agent-skills]], [[agent-evaluation]]. ^[../knowledge/raw/articles/blogwatcher/2026-07-08/2026-07-04-lil-log-harness-engineering-for-self-improvement.md] ^[../knowledge/raw/articles/blogwatcher/2026-07-08/2026-07-07-bair-blog-intelligence-is-free-now-what-data-systems-for-of-and-b.md] ^[../knowledge/raw/articles/blogwatcher/2026-07-08/2026-07-07-latent-space-ainews-the-field-guide-to-fable.md]