Deep research agent loops

Business constraint

An agent explores many documents under a time/token budget. Must not loop forever or thrash the same dead ends.

Naive v0

While true: retrieve → generate → repeat. Unbounded tool calls. Memory is a flat list.

Failure drill

Cost blowups. Duplicate fetches. Oscillation between two docs. No audit trail of why it stopped.

Evolution path

Iteration 1

Explicit state machine: plan → act → observe → critique → stop. Budgets on tokens, wall time, and tool calls.

Iteration 2

Memory store with ranked eviction (lab memory-rank). Idempotent tool results keyed by args hash.

Iteration 3

Backpressure: when tools are slow, shrink plan breadth. Persist run state for resume.

Implementation cut

Stop conditions are part of the product. Persist artifacts. Make runs replayable from the event log.

Numbers

Median tool calls/run. Duplicate tool rate. Budget exhaustion vs success rate.

Pattern tags

• backpressure
• outbox-idempotency

Self-check

1. Why unbounded loops fail economically?
2. What belongs in run state?
3. How does memory ranking help?
4. What is a good stop condition?
5. How do idempotent tools reduce thrash?

Walkthrough

Recordings will appear here when published. Until then, work the failure drill and lab locally — pause after each iteration and write the metric that moved.

Lab package: see /systems/labs/memory-rank/.