← Post-training series

Essay 06 / Long-horizon agents

Long-horizon agents need more than more context.

A larger window can hold more history. It does not automatically decide what matters, what changed, or what should be remembered next time.

SERIES 02 / ESSAY 06

06
LONG-HORIZON AGENTSHISTORY ≠ EXPERIENCE

KEEP → COMPRESS → UPDATE

  • Continuity
  • Priority
  • Recovery

A context window is a workspace, not a career. It can place yesterday's notes beside today's task, but it does not turn those notes into judgment or guarantee that the right lesson survives.

POST-TRAINING MODELSESSAY 06 / 10

More tokens delay forgetting. They do not solve it.

Larger context windows are genuinely useful. An agent can inspect more files, keep a longer conversation in view, and connect evidence that would otherwise be separated. But capacity is only the first constraint.

Long histories contain obsolete decisions, repeated facts, conflicting instructions, and details that matter only locally. An agent still has to select what deserves attention. Without prioritization, a larger window can become a larger pile: technically present, practically ignored.

Availability of history is not continuity of understanding.

A returning employee should outperform a new contractor.

Imagine two capable people joining a project. One is a contractor reading a carefully prepared handoff. The other has worked with the team for months. Both may possess the same documents, but the returning employee knows which conventions are load-bearing, which stakeholder worries are recurring, and which tempting solutions already failed.

Agents often behave like a sequence of contractors. Each task begins with another tour of the repository, another reading of the same notes, and another opportunity to repeat an old mistake. Long-horizon competence requires the equivalent of organizational experience: compressed, prioritized knowledge that changes future decisions.

Continuity should reduce:

  • Repeated questions about settled preferences.
  • Re-reading of unchanged foundational material.
  • Reintroduction of previously corrected mistakes.
  • Unnecessary escalation of familiar situations.

Remembering everything is not the objective.

A useful agent must decide what to retain, what to summarize, and what to discard. That decision depends on future value. A small user preference may be more important than pages of execution logs; a failed approach may matter precisely because it should not be repeated.

The knowledge also needs revision. Teams change priorities, systems change behavior, and users contradict earlier preferences. Durable memory without an update mechanism becomes a source of confident, stale errors.

This is why long-horizon design is inseparable from judgment. The agent needs a model of the user, the project, and its own uncertainty. It should know when earlier experience applies, when it conflicts with present evidence, and when clarification is worth the interruption.

Evaluate coherence across a sequence, not a single task.

Single-session success cannot show whether an agent becomes easier to work with. A better environment presents related work over time: recurring users, shared systems, changing requirements, and consequences that appear several tasks later.

  1. 01
    Preserve shared history.

    Let later tasks depend on preferences and decisions established earlier.

  2. 02
    Introduce change.

    Test whether the agent updates old beliefs when the environment moves.

  3. 03
    Measure repeated effort.

    Track redundant reading, questions, and failed approaches.

  4. 04
    Reward calibrated escalation.

    Knowing when not to act is part of durable competence.

Long-horizon ability is accumulated judgment.

More context provides room. The harder problem is turning experience into compact, revisable guidance that improves later action. Agents become durable collaborators when past work changes how they approach the next task, not merely how much text they can see.

View the complete series
Next · Essay 07

Continual Learning Is the Missing Layer

What it would mean for an agent to get better across related work.