Learn
What are long-horizon agents?
Long-horizon agents pursue goals that span hours, days, or weeks — planning, executing, checkpointing, and resuming across sessions. They need memory, tool access, evaluation, and governance beyond a single context window.
In depth
Short tasks fit in one prompt. Real company work — launch a campaign, close a quarter, migrate a service — spans many steps and interruptions. Long-horizon agents decompose goals, persist state, recover from failures, and report progress without the human re-prompting every step.
Building blocks: durable plan (tasks, dependencies), memory outside the window, tools with scoped permissions, checkpoints after major steps, evaluation that detects drift, and Ask on irreversible actions.
Reliability improved in 2025–2026 as models got better at tool use and platforms added orchestration — but long-horizon work still fails without explicit ownership and acceptance tests.
Neuro OS treats long jobs as roles with tasks and heartbeats — not one marathon chat session.
Examples
- Multi-day outbound campaign — research, draft, approve, send, follow up
- Quarterly content refresh — audit 200 pages, prioritize, ship updates
- Incident response — triage, draft postmortem, track action items over days
- Claude Code session + git — code agent across many files and commits
- Neuro OS task graph — role resumes from last checkpoint after restart
Related terms
FAQ
Why do long-horizon agents fail?
Context loss, missing checkpoints, no eval, unbounded tool access, and no human approval on high-stakes steps.
Do I need a bigger model?
Often you need better orchestration and memory — not only a larger window.
Run governed agent roles on a company OS — not only definitions in a glossary.