The one-sentence version
Long-horizon work takes many steps over an extended stretch — hours, days, or weeks — and success is only measurable at the end. The model is the engine; the scaffold around it is what gets the job across the finish line.
Short Horizon vs. Long Horizon
A short-horizon task is one prompt, one answer: "summarize this email." A long-horizon task is a project: dozens or hundreds of steps, where you only find out at the end whether the whole thing held together. Real work is overwhelmingly long-horizon:
Short horizon
- "Draft a reply to this email."
- "What does this clause mean?"
- "Convert this CSV to a table."
- Verified instantly, in a single glance
Long horizon
- "Migrate this mail system to M365."
- "Build, test, and deploy this workflow."
- "Reconcile these two systems and report the gaps."
- Success measurable only at the end
Why Long Tasks Fail: Three Compounding Problems
As tasks get longer, three failure modes stack up. None of them is "the model wasn't smart enough" — all three are structural:
1. Error compounding (drift)
Every step has some error rate. On a 5-step task, small mistakes get caught. On a 200-step task, even a 2% per-step error compounds toward near-certain failure — and an early mistake poisons everything downstream. The agent wanders off-goal without noticing.
2. Context exhaustion
The task outlives the context window. Step 180 needs a fact from step 3 — and the whiteboard is long gone. (See the Context Window section of Choosing a Model.) Long-horizon work needs external state: files, checklists, databases.
3. Credit assignment
The deliverable fails on Friday. Which of Monday's forty decisions caused it? Humans struggle with this too. Agents have no built-in mechanism for it — which is why verification has to be designed in, not hoped for.
What Actually Works: The Five-Pattern Scaffold
Long-horizon reliability comes mostly from architecture, not model size. The same five patterns show up in every serious agentic system:
Decompose. Turn the goal into an explicit plan of checkpoints before acting. "Migrate mail" becomes a numbered list of verifiable stages. The plan is the contract.
Checkpoint to durable state. Write progress to files, tables, or issue tickets — not chat history. A fresh context (even a brand-new session) can then resume exactly where things stopped.
Verify each checkpoint against the goal. Catch drift at step 20, not step 200. Verification reads real state — "show me the deployed file" — never the agent's claim that it's done.
Delegate slices. Subagents handle bounded chunks with fresh contexts; the orchestrator keeps only the thread. (This is the Workflow Architect pattern from Power User, applied across days.)
Re-plan after failure. A blocked step triggers a route change — new approach, updated plan — not a stall. Dead ends are information.
You already run this way: an agent keeping a plan file updated, ticking checkpoints, resuming after interruptions, and verifying against the real system before saying "done" — that's long-horizon agency in practice. The habits from the Playbook (decompose, verify, checkpoint) scale up; they don't change.
How "Horizon" Gets Measured
Researchers quantify this with the time horizon measure: how long a task a model completes with ~50% reliability. Two findings matter:
It's doubling fast. Measured horizons have been doubling roughly every ~7 months — from a few minutes of work (2023-era models) toward multi-hour tasks on frontier models.
Lab ≠ office. Benchmarks use clean, well-posed tasks (see why benchmarks lie a little, in Choosing a Model). Your production reliability at a given horizon will be lower — plan checkpoints accordingly. The metric is a trend indicator, not a procurement spec.
Why it matters for you: every model-generation jump extends the length of work a single delegation can carry. Tasks that needed a human-in-the-loop checkpoint last year may be one clean delegation this year — and the scaffold patterns above are what convert that raw capability into finished work.
Working Long-Horizon With Your AI
Practical translation for delegating a multi-day project to your agent:
Ask for the plan first. "Before you start, write out the stages and what 'done' looks like for each." Approve the plan, not just the outcome.
Insist on durable state. "Keep your progress in [file/table], updated at every step." Sessions can die; files don't.
Demand evidence at checkpoints. "Show me the output, then continue." — the Trust but Verify rule, applied at every stage instead of only the last.
Split when it stalls. If a step keeps failing, ask it to re-plan and delegate the slice — not to try the same thing harder.