CoreWise Academy

Prompting & Context · Layer II / Deep

Guardrails for hours-long agent runs

Guardrails are the support structure that keeps an hours-long agent run honest: an audit instruction, a memory that survives resets, and fresh-eyed verifiers.

Nº 016 · Vol. I·6 min read· Updated July 2026

Read firstStop shouting at the model (Covers softening trigger language and steering adaptive thinking. This guide takes those to autonomous, hours-long runs.) · The new Claude lineup and what it breaks (Assumes you know Fable 5's always-on thinking, refusal stop_reason, and effort model from the overview guide.)
“a model that runs for hours needs a support structure”

Claude Fable 5 is built to take multi-minute turns and run autonomously for hours. That is the capability you are paying for, and it is the one most setups quietly waste. Point a model built for the long run at a prompt written for quick turns and you get the worst of it: an agent that works for 2 hours, then hands you a confident summary of things it never actually did. The fix is the support structure that keeps a long run honest, and a first move most people miss because it is a single dial.

01Effort is the dial, and it moved

On Fable 5 the effort parameter is the primary control over intelligence, latency, and cost. High is the default and the right resting place for most work. Push to xhigh when the task is capability-sensitive, the kind of problem where getting it right matters more than getting it fast. Drop to medium or low for routine work. Low effort on Fable 5 can still exceed xhigh on prior models. So “cheap and fast” no longer means “dumb.” The reflex to reach for a bigger model on a hard task is often better answered by keeping the same model and raising the effort, and the reflex to save money by downgrading the model is often better answered by dropping effort on Fable 5 instead.

EffortReach for it whenWhat to expect
high, the defaultMost work; the right resting placeStandard agentic turns
xhighCapability-sensitive problems where right beats fastTurns of 2 to 10 minutes, and worth it
medium / lowRoutine tasks that still need Fable-level reasoningThe cheapest runs; low can still exceed xhigh on prior models
effort stops paying off, see Rent the model, own the method.

02Make the long run check itself

The signature failure of a long autonomous run is the fabricated status report: the agent claims it ran the tests, fixed the file, deployed the change, when it did none of those. The best-practices guidance names the fix, and it is one instruction. Tell the model to audit every claim against tool results before reporting it. In Anthropic’s own testing that instruction nearly eliminated fabricated status. It works because it removes the room to guess.

“A claim now has to trace to an observed tool result or it does not get made.”

The countdown is the other trap. A running remaining-token counter reads to the model as a deadline, and it responds by wrapping up early, cutting work short to “finish in time.” So do not show remaining-token counts. If your software must show them, pair the number with an explicit reassurance that the model has ample context left, so the count informs without making it wrap up early. And set your software’s timeouts for the runs you actually expect. Turns of 2 minutes or more are normal here, not a hang.

03Memory and fresh eyes

A run longer than one context window needs saved information that survives the reset. Give the model a memory system of Markdown files, one lesson per file, so a durable finding written 2 hours in is still readable 4 hours in. Small focused files beat one growing scratchpad, the same discipline the managed-agent memory docs recommend for their stores. For the managed-agent version, stores, scopes, and caps included, see Give your agent a memory.

For checking work, prefer a fresh-context verifier subagent over self-critique.

“A model reviewing its own transcript shares its blind spots; a subagent that reads only the finished work and the criteria does not.” Keeping the evaluator separate from the worker is the same split behind Automate the step that slows you down. And if you want the model to send you a word-for-word message mid-task, define a send_to_user tool in your own software, then prompt for it explicitly. Defining the tool alone is not enough. Fable 5 needs the nudge, a clear instruction on when and how to call it, or it stays silent.

You cannot build your way into reading the raw chain of thought. On Fable 5, thinking.display defaults to omitted and the raw reasoning is never returned. Worse, an instruction telling the model to echo or transcribe its reasoning can trip the reasoning_extraction refusal, returning stop_reason “refusal” instead of work. So audit your skills for show-your-thinking language written for older models, strip it, and set thinking.display to summarized to read structured thinking blocks instead.

Further reading

FAQ

Why does surfacing a running token countdown make a long-running agent finish worse?

The countdown reads as a deadline. Shown how few tokens remain, the model wraps up early, cutting corners on work it would otherwise finish. Anthropic's guidance is to not show remaining-token counts at all, and if your software must show them, to reassure the model it has ample context so the number does not make it wrap up early.

What single instruction most reduced fabricated status reports in Anthropic's testing, and what does it actually change?

An instruction to audit every claim against tool results before reporting it. In testing it nearly eliminated invented status updates. It works because the failure was the model narrating success it never confirmed. Forcing each claim back to an observed tool result removes the room to guess.

Why can't you prompt Fable 5 to print its reasoning, and what do you read instead?

Instructions telling the model to echo or transcribe its internal reasoning can trip the reasoning_extraction refusal category, and the raw chain of thought is never returned regardless. thinking.display defaults to omitted; set it to summarized to get readable thinking summaries, and read those structured blocks instead of asking the model to show its work in prose.

For your agent

A distilled, agent-ready version of this guide, not a skill. Copy it below, or point the agent at corewise.academy/guides/guardrails-for-long-runs.md.

file · guide.md guardrails-for-long-runs 51 lines · markdown
# Guardrails for hours-long agent runs Consult when configuring an agent (Fable 5) to run autonomously for many minutes or hours, and it hands back confident summaries of work it never did, wraps up early, or forgets findings across a context reset. Distilled from the CoreWise guide `guardrails-for-long-runs`, whose sources are Anthropic's prompting-Fable-5, best-practices, and managed-agent memory docs. For live model ids, pricing, and exact refusal or 400 errors, use the `claude-api` skill rather than trusting version-pinned facts from memory. ## Set effort, not model size The effort parameter is the primary dial for intelligence, latency, and cost on Fable 5. - high is the default and the right resting place for most work.- xhigh for capability-sensitive work where getting it right beats getting it fast. Expect turns of 2 to 10 minutes.- medium or low for routine tasks. Low on Fable 5 can still exceed xhigh on prior models, so cheap and fast is not dumb. Rule of thumb: a hard task usually wants the same model at higher effort, not a bigger model. Saving money usually wants lower effort on Fable 5, not a downgraded model. ## Make the run check itself 

+ 35 more lines · the copy includes all of them

then save as .claude/reference/guardrails-for-long-runs.md
Sources

Original work by the editor, no video source.