Skip to content

PATTERN Cited by 1 source

Fresh-session agent iterations

Pattern

Create a new agent session for every work iteration and reconstruct its working context from explicit durable state: the current goal, selected peer traces, and artifacts. Do not carry an unbounded private conversation history across iterations. (Source: sources/2026-08-11-aws-scaling-patterns-for-self-organizing-multi-agent-clusters-with-kiro)

Problem

A persistent agent session can retain an early, narrow interpretation of the task after peers have found better information. That accumulated conversational history creates behavioral momentum: the agent continues optimizing its previous plan rather than re-evaluating the current shared environment. In a large independent-agent cluster, the effect is drift rather than useful continuity.

Mechanism

  1. Start a fresh session.
  2. Read the durable goal and a bounded peer/artifact view.
  3. Choose and perform one contribution.
  4. Persist the result and next intent in shared state.
  5. Discard the session history; repeat from the current external state.

The technique relies on the model's fundamental statelessness: conversational continuity is supplied by caller-provided context, so deliberately selecting the next context is a control surface. See concepts/agent-loop-stateless-llm.

Trade-offs

  • Gain: reduces drift and makes the visible shared record the source of coordination truth.
  • Gain: a crashed process loses no hidden coordination state needed by the next iteration.
  • Cost: repeatedly loading context consumes tokens and can omit useful private reasoning.
  • Cost: stale shared files become more dangerous, so runs need explicit archival or namespacing.

Seen in

Last updated · 619 distilled / 1,953 read