Agent Goals Are Becoming Runtime State

The next agent reliability problem is deceptively simple: when an agent says it has a goal, where does that goal live?

If the answer is “somewhere in the conversation,” the product eventually runs into edge cases. A session that starts with a goal may need to appear in a resume list before any normal user message exists. A long pasted task brief may need to become a concise stored objective. A follow-up task may need to wait behind the current one. A restarted agent may need to know what target it was still pursuing.

This month’s public evidence from OpenAI Codex and Hermes Agent points to the same pressure from different directions. The goal is becoming a typed piece of runtime state.

Codex separates the goal from the first message

OpenAI Codex PR #21981, merged on May 11, is a small metadata change with a large design signal. It fixes /goal-first threads that could be missing from codex resume and app recents because discovery had depended on metadata derived from a normal first user message.

The earlier PR #21489 tried a narrower visible fix: persist ThreadGoalUpdated events and synthesize first_user_message from the goal objective. The review discussion pushed back on that shape. A goal objective can provide useful preview text, but it is not literally the first user message, and reusing that field would blur title, preview, and discovery behavior.

The merged replacement adds an internal thread preview metadata field. ThreadGoalUpdated can now be preview-bearing metadata, while first_user_message keeps its original meaning. That is the story: Codex did not just make a list item appear. It gave goal-derived display state a separate lane.

A nearby Codex issue, #21477, shows the same boundary at input time. The report says a long pasted /goal command was rejected before the model could normalize it into a concise persisted ThreadGoal.objective. The issue argues that raw goal source text and stored thread objective should not be treated as the same object.

Hermes makes goals persistent work control

Hermes Agent exposes the broader product version of the same idea. Its v0.13.0 release, published May 7, describes /goal as a primitive that keeps the agent locked on a target across turns. In the same release, ACP /steer and /queue let editor integrations direct the in-flight agent or queue follow-up work.

The project’s release file places persistent goals beside checkpoints, session durability, atomic session persistence, and reasoning-metadata preservation. That grouping matters. A goal that survives across turns is no longer just a slash command; it belongs to the same layer as restart behavior and session storage.

Hermes issue #22617, opened two days after that release, pushes the idea further. The request asks for /queue-goal <text>, automatic promotion when the current goal completes, persistence across session restarts, and pending items visible in /goal status. That is not shipped behavior in the issue. It is evidence of what users expect once goals become durable: ordering, status, and continuity.

The implication

The shared pattern is not that Codex and Hermes have the same design. They do not. Codex is carving goal-derived preview out of normal user-message metadata. Hermes is presenting goals as a persistent operating target and connecting nearby queue and durability surfaces. But both are answering the same product question: how does an agent runtime remember what it is trying to finish without hiding that memory inside ordinary chat text?

That question gets sharper as agents move from one-turn assistance to longer tasks. Goals need names. They need limits. They need history. They may need queues. They may need to survive process restarts. And when they affect what the user sees in a resume list or editor panel, they need their own metadata rather than borrowing fields that meant something else.

Watch next for goal inspectors, queued objective lists, restart-safe goal histories, and clearer boundaries between “what the user said” and “what the runtime is currently pursuing.” The durable agent is not just the one that remembers more context. It is the one that can account for its objective as state.