Skip to content

Guided tour

FabrCore 2.0 · Release and package availability

These guides track the current 2.0 source. Stable 2.0.0 publication is pending; package commands show the release target. Until it is published, follow the source quick start or use a matching available prerelease set. Release migration · Runtime modes

MODULE 06 · LESSON 6.2

Plan, execute, and bound the loop

Control planning, execution and loop budgets through agent arguments.

Lesson 34 of 86 · FabrCore 2.0

Overview

Planning mode gives you a readable plan. Execute mode performs work. These modes do not implement a durable human approval transaction by themselves. A bounded loop can stop with work remaining; the application must make that state visible.

Planning and execution are explicit modes

Planning describes intended work; execution performs the permitted actions. A mode value is not a durable approval record from a human. If a business write requires review, store the proposed action, owner, expiry and digest in application state and validate a later decision before applying it. A prompt saying “approved” is not that validation.

There are two different iteration limits

The outer harness loop can repeat model runs until its completion evaluator is satisfied or its budget is exhausted. A single model run can itself invoke tools repeatedly, controlled by its per-request limit. Lowering the outer count does not necessarily bound tool calls within one run. Choose modest limits while learning and inspect why a run stopped.

Set a small execution budget

  1. Set _HarnessDefaultMode, _HarnessTodo and _HarnessLoop in AgentConfiguration.Args. Choose a small iteration budget for the first multi-step exercise.
  2. Use todo/background completion evaluators only when their providers/delegates exist. Set per-request tool iteration and timeout bounds separately from the outer loop budget.
  3. For a human-approved mutation, persist an application-owned request/digest/owner/expiry, end the turn and validate a later decision before consuming approval and performing the effect.
AgentConfiguration.Args · small learning budget 
{
  "Args": {
    "_HarnessDefaultMode": "plan",
    "_HarnessTodo": "true",
    "_HarnessLoop": "todo",
    "_HarnessLoopMaxIterations": "2",
    "_HarnessMaxIterationsPerRequest": "5"
  }
}

Force a visible budget stop

  1. Use the Args fragment below on a disposable harness instance and ask for a task with several distinct fixture steps. First inspect the plan mode's proposed work.
  2. Use the documented execute path and observe todos, calls and final status. With a deliberately small budget, unfinished work should remain visible when the loop stops.
  3. Increase only the outer iteration budget through explicit reconfiguration and repeat the same task. Compare completed steps and call count; do not treat a longer answer as proof of more completed work.

The cap is a stopping condition, not a completion condition. A correct response explains which steps finished and which require another decision or run.

If the result is different

General durable channel tool approval is not a built-in harness feature. A model switching its own mode is not a trusted human approval.

Go deeper

Explore the related documentation.