Skip to content

Latest commit

 

History

History
111 lines (79 loc) · 2.12 KB

File metadata and controls

111 lines (79 loc) · 2.12 KB

Context

Context engineering is deciding what the agent sees now, what it can retrieve, and what must be stored durably.

More context is not always better. The right context is scoped, current, source-backed, and relevant to the worker.

Context Layers

Immediate Context

What the agent needs for the next step:

  • active goal
  • typed intent
  • current task
  • relevant constraints
  • relevant policy
  • last few state transitions

Working Context

What the room is currently managing:

  • goal graph
  • task graph
  • active workers
  • pending approvals
  • current artifacts
  • open risks

Durable Context

What should survive the session:

  • artifacts
  • beliefs
  • decisions
  • user preferences
  • eval results
  • failure modes

External Context

What must be fetched because it changes:

  • news
  • prices
  • laws
  • product docs
  • API docs
  • schedules
  • model availability
  • deployment state

Worker Context Packs

Each worker should receive a context pack, not the entire transcript.

Example:

# Worker Context

Goal: ...
Task: ...
Constraints:
- ...

Policy:
- web research: allowed
- external actions: blocked
- budget remaining: 12 workers

Relevant artifacts:
- ...

Relevant beliefs:
- ...

What Not To Include By Default

  • full transcript
  • unrelated goals
  • stale artifacts
  • private data not needed for the task
  • tool credentials
  • hidden system state the worker cannot use
  • speculative beliefs without source labels

Context Freshness

Each claim should be classified:

  • stable: unlikely to change
  • dynamic: may change and should be checked
  • current: verified recently
  • stale: should not be used without refresh
  • user-provided: true within the room unless corrected

Context Failure Modes

  • Lost steer: user intent was spoken but never stored.
  • Context flooding: too much transcript causes the model to miss the actual task.
  • Stale authority: old worker output overwrites a newer goal.
  • Hidden policy: model attempts a tool the room would not permit.
  • Fake memory: model remembers a fact that was never committed.

Rule

Context should be sufficient to do the work and narrow enough to make mistakes observable.