Articles/Context engineering
Context Engineering Turns One: The Four Techniques, a Year Later
Anthropic’s context engineering post named four techniques in September 2025. A year on, each has become a product feature. Here is where each one stands and what is still missing.
Jose R. Gonzalez, CPA · Founder of ContxOS
Updated · 2 min read
In short
On September 29, 2025, Anthropic’s "Effective context engineering for AI agents" named four techniques for working within a model’s attention budget: just-in-time retrieval, compaction, structured note-taking and sub-agent architectures. A year later every one of them is a shipped feature. Tool search and MCP bring just-in-time retrieval, compaction is built into the Claude API, notes have become memory files and memory stores, and subagents are standard. What is still hard is the note-taking layer: keeping notes structured, current and shared once several agents write them over months.
The four techniques
- Just-in-time retrieval: keep references in context and load the content only when it is needed.
- Compaction: summarize a long conversation so it can continue.
- Structured note-taking: have the agent write notes outside the window (the post used a NOTES.md file as its example) and read them back later.
- Sub-agent architectures: send focused work to agents with clean contexts that return condensed results.
Where each one is now
- Retrieval: MCP is the standard way to reach tools and data. Its July 2026 revision and roadmap focus on scaling to large tool catalogs through progressive discovery.
- Compaction: server-side compaction in the Claude API (February 2026), on-demand compaction (September), and PreCompact hooks in Claude Code.
- Notes: CLAUDE.md and AGENTS.md for rules, Claude Code auto memory, Managed Agents memory stores and Letta’s memory files for everything else.
- Sub-agents: Claude Code Agent Teams, Managed Agents multiagent orchestration and Codex subagents all shipped this year.
The gap: notes at scale
Structured note-taking works well for one agent on one task. It gets harder once several agents write notes over months. Notes contradict each other, old ones stay after they stop being true, and a flat file cannot say which decision replaced which. Claude Code added a /doctor prompt-audit check in September for stale or conflicting instructions, which shows the problem is real.
The LongMemEval-V2 benchmark from May 2026 points the same way. Across very long agent histories, a plain retrieval baseline scored 48.5%, well below methods that organized memory.
What it means in practice
Treat notes as a data model, not a text file: typed entries, explicit links between versions, freshness on every read, and a shared home that every agent and client can reach. That is the layer ContxOS provides, next to the rules files and compaction your tools already have.
Sources
- Effective context engineering for AI agents — Anthropic Engineering, September 29, 2025
- MCP 2026-07-28 specification release — Model Context Protocol blog, July 28, 2026
- MCP roadmap — Model Context Protocol blog, August 22, 2026
- Claude Platform release notes — Anthropic
- Claude Code changelog — Anthropic on GitHub
- LongMemEval-V2 — arXiv 2605.12493, May 12, 2026
Frequently asked questions
Is context engineering the same as prompt engineering?
It is broader. Prompt engineering is about the instructions. Context engineering is about everything that enters the window: tools, retrieved data, notes and history.
Give your agent a memory that lasts
Connect ContxOS to Claude in about a minute. The free plan includes one project.
Product names belong to their owners. ContxOS is not affiliated with the companies named here.