Agent Runtimes Are Learning When To Stop

Reporter Notes

The strongest shared pattern is not "more controls" in the abstract. It is that agent systems are making the end of a run and the continuity between runs inspectable:

• OpenCode makes a configured step limit ask for a summary with tools disabled.
• Codex exposes lineage identifiers for context windows across compaction, resume, and rollback.
• OpenClaw sends Copilot runtime events through harness hooks.
• Hermes offers an in-place compaction path that keeps one session ID.

The under-covered source audit named OpenCode. That repo produced the lead's freshest concrete fact and the people companion, so the coverage gate is satisfied without forcing an unrelated source into the story.

The nearest repetition risk is June 14's recovery-path article and June 19's terminal control panel article. The new material is the stop-state contract: what happens when the runtime reaches a boundary, not simply what the user can see or click.

Primary Evidence

• OpenCode commit 4f1a9d7a: https://github.com/anomalyco/opencode/commit/4f1a9d7aef56163d8fe265f7f9cbe295cc1df95a
  • Evidence used: Kit Langton's patch changes configured agent step-limit behavior, adds a core max-step prompt, and updates session-runner tests.
• OpenCode max-step prompt file: https://github.com/anomalyco/opencode/blob/4f1a9d7aef56163d8fe265f7f9cbe295cc1df95a/packages/core/src/session/runner/max-steps.ts
  • Evidence used: the prompt instructs a capped agent to stop using tools and summarize completed and remaining work.
• OpenCode session-runner test file: https://github.com/anomalyco/opencode/blob/4f1a9d7aef56163d8fe265f7f9cbe295cc1df95a/packages/core/test/session-runner.test.ts
  • Evidence used: tests assert the final limited turn has no tools available and that uncapped agents can continue beyond the old local tool-step cap.
• Codex context-window lineage commit: https://github.com/openai/codex/commit/d1209bddfce79e8f6cafd7f9212a2df6915aec5d
  • Evidence used: Codex adds first, previous, and current context-window IDs and carries them through compaction and reconstruction paths.
• OpenClaw Copilot lifecycle hooks commit: https://github.com/openclaw/openclaw/commit/03b022b88e7586260225b8b20b8ea53c18f7d40f
  • Evidence used: OpenClaw routes Copilot prompt, tool, output, and agent-end events through harness lifecycle hooks.
• Hermes Agent in-place compaction commit: https://github.com/NousResearch/hermes-agent/commit/47fadc24d79c1ff21b23518c0e27aaa3146a421d
  • Evidence used: Hermes adds an option to compact a conversation while keeping the same session ID.

Evidence Limits

• These sources do not show a shared standard or coordinated design across projects.
• The article does not claim the patches are released to every end user.
• The Codex and OpenClaw evidence is implementation-level; the Hermes and OpenCode evidence shows options and tests, not universal default behavior across all deployments.