Agentic demo

The Long-Horizon Agent Run

Give Claude Fable 5.1 a task too long for one turn and ask it to plan, execute, checkpoint its own progress, and report what it verified.

Cast: Fable 5.1Engineering leads, ops teams, automation ownersAdvanced45-120 minutes

"adaptive thinking"

Casting: which model to run

This is the demo Claude Fable 5.1 exists for. Anthropic positions it for demanding reasoning and long-horizon agentic work, with always-on adaptive thinking and a default effort level of high inside Claude Code. Running this on a faster model is a different demo with a different result.

See the casting page for the full current lineup and API IDs. MagicClaude does not score models; Claude Reports covers benchmark comparisons and pricing analysis.

Setup: what the audience sees

Most Claude demos fit in one answer. This one does not. The reveal is coming back later to a plan file with items crossed off, a set of small commits, and a report that distinguishes what the model verified from what it merely changed.

The setup matters because Claude demos fail when the prompt hides the goal, mixes private data into a public reveal, or asks for a finished answer without giving the model a way to expose assumptions.

  • Use Claude Code on a throwaway branch of a repo you own, or a sandboxed project directory.
  • Pick a task with many small verifiable steps rather than one hard step: a migration, a test-coverage sweep, a dependency audit, or a docs-to-code consistency pass.
  • Write the acceptance criteria down before you start, because the whole point is checking the run against something you wrote first.
  • Give Claude a scratch plan file it is allowed to rewrite, so progress survives an interruption.

What this does not prove

It does not prove the run was correct, and it is not a measurement. A long autonomous run produces more code to review, not less. If you want scored comparisons between Fable 5.1 and Opus 5 on agentic benchmarks, that analysis belongs on Claude Reports, not here.

Copyable promptPlain text
<task>
Work through the checklist in PLAN.md until every item is either done or blocked.
</task>
<operating_rules>
- Before starting, restate the acceptance criteria in your own words and stop if they are ambiguous.
- Work one checklist item at a time. Update PLAN.md after each item with status and a one-line note.
- Run the narrowest useful verification command after each item and record the command you ran.
- Commit after each completed item with a message naming the item.
- If an item is blocked, mark it BLOCKED with the reason and move on. Do not invent a workaround.
- Do not touch files outside the directories listed in PLAN.md.
</operating_rules>
<final_report>
When the checklist is exhausted, produce:
1. Items completed, with the verification command and its result for each
2. Items blocked, with the reason
3. Changes you made but did not verify
4. What a human should review first
</final_report>

The reveal: what should happen

  • Claude should restate the acceptance criteria before it starts working, which is where most long runs go wrong.
  • PLAN.md should read like a work log afterward, with per-item status you can diff against what you wrote.
  • The final report should separate verified work from unverified work instead of declaring the whole run a success.
  • A blocked item marked honestly is a better outcome than a clever workaround, and the prompt is written to prefer it.

Verification steps

A demo becomes useful when the audience can inspect it. Use the steps below to turn the reveal from theater into a repeatable workflow.

  1. Read the commit sequence, not just the final diff, and confirm each commit maps to a checklist item.
  2. Re-run the verification commands the report claims to have run and compare the output.
  3. Check the "changed but not verified" section against the diff, because anything missing from that list is an overclaim.
  4. Confirm no files outside the declared directories were touched.

Guardrails

  • Use a throwaway branch and no production credentials. A longer run means more time for an unsupervised mistake to compound.
  • Keep the task list narrow and explicit. Open-ended instructions like "improve the codebase" remove your ability to check the result.
  • Treat the run as unreviewed work until a human reads the diff, regardless of how confident the final report sounds.
  • Long agentic runs consume real usage. Check your plan limits before starting a demo you intend to let run.

Variations

  • A dependency audit that produces a report and no code changes at all.
  • A test-coverage sweep limited to one module, with a coverage number before and after.
  • A docs-versus-code consistency pass that only files issues rather than editing code.

FAQ

Why Fable 5.1 instead of Opus 5 for this one?

Anthropic positions Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, and recommends starting with Claude Opus 5 otherwise. This demo is the case the guidance describes, and it is the only demo in this gallery where the model choice changes the shape of the result rather than just the speed.

Is this a benchmark of agentic ability?

No. One run on one repo tells you nothing generalizable. It is a demo of a supervision pattern: written acceptance criteria, a checkpointed plan, per-item verification, and an honest final report.

What if the run stops partway through?

That is why PLAN.md exists. A run that dies at item six should be resumable from item six, and a demo that cannot be resumed is a demo that hid its state inside the conversation.

Keep the act moving

Related demos and tools

Claude Code

Codebase Debugging

Use Claude Code to read a repo, reproduce a bug, make the narrow fix, and report the tests it ran.

Model behavior

Effort Dial

Run one ambiguous task at low and high thinking effort, then read the two answers side by side to show what deliberation actually buys.

Tool

Demo Gallery Tool

Find more exact prompts by format, audience, and difficulty.

Tool

Prompt Remixer

Turn this prompt pattern into a themed copy-paste variant.

Primary Sources

These demos cite official Anthropic docs or help-center pages for capability claims. They do not claim identical output across accounts, plans, models, or dates.

Claude Platform Docs

Claude Fable 5.1 overview

Claude Fable 5.1 (`claude-fable-5-1`) has a 1M-token context window, 128K max output, always-on adaptive thinking, and a default effort level of high in Claude Code and medium on Claude.ai and Cowork.

Anthropic

Claude Fable 5.1 and Mythos 5.1

The September 1, 2026 announcement introduces Fable 5.1 as Anthropic’s most capable generally available model and describes cheaper cache reads for agentic workloads.

Claude Code Docs

Claude Code overview

Claude Code is an agentic coding tool available in terminal, IDE, desktop app, and browser surfaces.

Claude Platform Docs

Code execution tool

The code execution tool runs Bash and Python in a sandbox to analyze data, generate files, and manipulate files.

Claude Platform Docs

Choosing a model

Anthropic guidance is to start with Claude Opus 5 for most work, move to Claude Fable 5.1 for demanding reasoning and long-horizon agentic tasks, use Sonnet 5 for the speed/intelligence balance, and Haiku 4.5 for the fastest and cheapest runs.