Context management

ACP context management spans more than prompt text. It includes thread history, project scope, computer state, attached files, selected agent behavior, and the decisions you make about continuation versus reset.

Control what the agent sees, what state it inherits, and what should persist between runs.

Think in context layers

ACP context comes from multiple layers at once: project, thread, computer, agent, files, and runtime state. Good context management means being explicit about which of those layers a run should inherit.

Treat context as a stack, not just a prompt window.

Be explicit about whether work should inherit prior history.

Use projects, threads, agents, and computers together intentionally.

Control context size and relevance

The best ACP workflows give agents enough state to act without drowning them in stale or irrelevant history. That means deciding when to continue, when to fork, and when to start a clean thread or computer.

Continue when prior state is truly relevant to the next task.

Fork when you need to branch without losing prior work.

Reset when the old state is more confusing than useful.

Use context as an operational primitive

Context management in ACP is part of system design. It influences cost, execution quality, safety, and how easily work can be resumed or audited later.

Model context choices as part of application architecture.

Minimize accidental carry-over from unrelated work.

Use explicit ACP primitives instead of stuffing everything into prompts.