Evidence Trail

Agent Goals Are Becoming Runtime State

May 21, 2026 / Daily Edition / 7 source signals.

repo openai/codex main
7 source signals 2 repos f10ddc3
> f10ddc3 / May 21, 2026 / Daily Edition
Read Story Open Edition

Reporter Notes

The story is about a boundary becoming explicit. A chat command like /goal looks small in a UI, but Codex and Hermes both expose the runtime consequences: goals affect session discovery, preview metadata, persistence, queued follow-up work, restart behavior, and user expectations about chained objectives.

Codex is the cleaner implementation case because PR #21981 separates preview metadata from first_user_message. The review path from #21489 matters because it shows the design tension: making a goal masquerade as the first user message solves listing, but muddles semantics. The merged PR turns that into a separate metadata lane.

Hermes is the broader product case. v0.13.0 places /goal beside checkpoints, auto-resume, ACP /queue, and reasoning metadata preservation. Issue #22617 shows that once a goal persists, users naturally ask for multiple goals, promotion, status, and restart durability.

Primary Evidence

  • OpenAI Codex PR #21981, Use goal preview metadata for goal-first threads: https://github.com/openai/codex/pull/21981
  • Evidence used: merged on May 11, 2026; adds internal thread preview metadata; treats ThreadGoalUpdated as preview-bearing metadata; keeps first_user_message distinct; merge commit f10ddc3.
  • OpenAI Codex PR #21489, Show goal-started threads in resume lists: https://github.com/openai/codex/pull/21489
  • Evidence used: opened on May 7, 2026; earlier approach persisted ThreadGoalUpdated events and synthesized first_user_message from the goal objective; review feedback questioned the shape.
  • OpenAI Codex issue #21477, /goal rejects long pasted input before Codex can normalize the objective: https://github.com/openai/codex/issues/21477
  • Evidence used: the issue distinguishes raw /goal source text from a bounded persisted ThreadGoal.objective.
  • Hermes Agent releases page: https://github.com/NousResearch/hermes-agent/releases
  • Evidence used: v0.13.0 release on May 7, 2026 says /goal keeps the agent locked on a target across turns and ACP /steer and /queue direct in-flight or follow-up work. v0.14.0 on May 16, 2026 establishes current project activity.
  • Hermes Agent RELEASE_v0.13.0.md: https://github.com/NousResearch/hermes-agent/blob/main/RELEASE_v0.13.0.md
  • Evidence used: persistent /goal PRs #18262, #18275, and #21287; ACP /steer and /queue PR #18114; session persistence PRs #20279 and #20296; checkpoints v2 PR #20709.
  • Hermes Agent issue #22617, [Feature]: Support for queueing goals (goal queue): https://github.com/NousResearch/hermes-agent/issues/22617
  • Evidence used: proposed /queue-goal <text>, automatic promotion after current goal completion, persistence across session restarts, and /goal status.

Evidence Limits

  • This article does not claim Codex and Hermes share a goal implementation or standard.
  • Codex review discussion described goals as experimental, so the article treats the metadata design as current evidence rather than final architecture.
  • Hermes issue #22617 is a feature request. It is evidence of user/operator pressure, not proof that /queue-goal has shipped.

Open Questions

  • Codex goals were described as experimental in review discussion, so the article should not claim the shape is final.
  • Hermes issue #22617 is a feature request, not an accepted or merged implementation.
  • The accessible evidence supports convergence around durable objective state, but not a shared standard or shared implementation between projects.