NEW RUNTIME

Topic hub

Context engineering

How teams package memory, rules, traces, tools, and task state so agents can work without restarting from raw chat context.

Short answer

  • Context engineering is the work of turning scattered organizational knowledge into durable inputs an agent can inspect and act on.
  • The strongest systems separate reusable rules, task state, source evidence, and operator corrections instead of burying everything in one prompt.
  • The practical payoff is lower context waste and more reproducible agent behavior across long-running tasks.

Pattern memory

Current hypotheses

1 patterns
  1. medium

    Skills become a portable capability layer

    Agent skills are emerging as a portable capability layer, but their value depends on progressive disclosure, provenance, security review, and behavioral evaluation.

Field notes

Longer analysis

3 notes

  1. AI-Native Engineering Teams Are Smaller and Stricter

    AI-native engineering changes team size, roles, onboarding, metrics, and review: code is no longer written only by humans, but context and quality ownership get stricter.

  2. Claude Code Subagents Split Memory Instead of Inflating One Session

    Claude Code subagents show a practical form of agent memory hygiene: move noisy work into a separate context and return only the compressed result to the main session.

  3. Unabyss Builds a Shared Context Layer Across Claude, Codex, and Cursor

    Unabyss points to where agent tooling is going: context stops being one client's internal memory and becomes a separate product layer with governance and portability.

Raw signals

Recent observations

9 signals

  1. Agent skills need behavioral evals, not prose review

    Benchmarks show that expert-authored skills can help while self-generated skills can underperform a no-skill baseline.

  2. Retrieval and long context do not create company memory

    Finding a relevant fragment does not establish whether it is current, authorized, trustworthy, or applicable to the present task.

  3. Claude Code Startup Context Can Be Reduced to 13K Tokens

    A measured context audit shows that tool descriptions and default instructions consume a material share of the initial window before repository work begins.

  4. Codebase Memory Combines Search with a Repository Map

    A code retrieval tool adds structural repository context to text search so an agent can locate relevant code without loading the whole project into context.

  5. Pydantic AI loads capabilities only when needed

    A capability can bundle instructions, tools, model settings, and hooks while exposing only a compact description until activation.

  6. CLAUDE.md Shrinks as Skills, Hooks, and Subagents Mature

    Agent behavior moves out of one giant context file into scoped mechanisms that load only for the relevant task and lifecycle stage.

  7. Production AI Engineering Is the System Around the Model

    Examples from large engineering teams point to testing, context layers, security, and workflow integration as the durable production work.

  8. Codex Subagents Reduce Context Pollution

    Delegating bounded work to subagents preserves the parent agent's attention and creates clearer evidence boundaries than loading every exploration step into one conversation.

  9. SKILL.md Acts as a Behavior Loader, Not a Better Prompt

    Skills package procedures, tools, and progressive context so an agent can load behavior only when the task requires it.