Skip to content

loop.md

Latest commit

 

History

History
152 lines (106 loc) · 2.6 KB

loop.md

File metadata and controls

152 lines (106 loc) · 2.6 KB
Raw

Loop Engineering

Agent systems are made of loops. Reliable systems name the loops, bound them, and make each transition visible.

Primary Loop

flowchart LR
  observe["Observe"] --> interpret["Interpret"]
  interpret --> plan["Plan"]
  plan --> act["Act"]
  act --> verify["Verify"]
  verify --> remember["Remember"]
  remember --> observe
Loading

Loop Steps

1. Observe

Collect current state:

  • human input
  • room state
  • active goal graph
  • policy
  • worker statuses
  • recent artifacts
  • relevant beliefs
  • tool results

Observation must be bounded. The agent should not ingest the whole world by default.

2. Interpret

Classify the new input:

  • new goal
  • add parallel goal
  • replace goal
  • constraint
  • correction
  • approval
  • question
  • cancellation
  • retry request
  • status request

Interpretation should produce typed intent, not only prose.

3. Plan

Convert intent into state transitions:

  • create goals
  • split tasks
  • choose skills
  • assign workers
  • set dependencies
  • reserve budget
  • request permission if needed

The plan is a proposal until committed by the harness.

4. Act

Run allowed actions:

  • deterministic reducer step
  • model call
  • tool call
  • worker schedule
  • artifact write
  • trace write

Actions should be idempotent where possible and guarded against stale commits.

5. Verify

Check whether the action achieved the intended state:

  • schema validation
  • artifact exists
  • source list exists
  • tests pass
  • browser flow works
  • logs are clean
  • user-visible state updated

A model saying "done" is not verification.

6. Remember

Write durable memory only after verification:

  • artifact
  • belief
  • decision
  • failure mode
  • eval case
  • user preference

Memory should keep source, confidence, and timestamp.

Loop Types

Foreground Loop

Handles live conversation and turn-taking.

Intent Loop

Turns human steering into typed state transitions.

Worker Loop

Moves work from queued to running to terminal state.

Verification Loop

Checks outputs across the real boundary the user cares about.

Memory Loop

Consolidates durable lessons from artifacts, traces, and corrections.

Loop Control

Every loop needs:

  • stop condition
  • max attempts
  • budget
  • trace events
  • retry policy
  • stale write guard
  • user-visible state

Bad Loop Smells

  • unbounded "think more" calls
  • transcript-only memory
  • repeated acknowledgements
  • silent retries
  • workers that cannot be canceled
  • stale workers that still commit
  • no visible blocked state
  • no external verification

Rule

Loops should continue because state requires it, not because the model keeps talking.