planning loops

Linear plans fail the moment they hit an unexpected dependency or a hidden constraint. This skill moves your agent beyond static to-do lists, implementing adaptive planning loops that treat execution as a source of new information. It ensures your agent doesn't just finish a list of steps, but actually reaches the goal by replanning when reality deviates from the initial hypothesis.

What it does

• Plan-Execute-Replan updates the remaining steps whenever execution results invalidate the original strategy.
• Dynamic Workflow enables state-machine logic where the next step is chosen based on the immediate output of the previous action.
• Goal Decomposition recursively breaks large, complex objectives into directly executable subgoals before execution begins.
• Progress Evaluation acts as an objective sanity check to measure actual goal proximity, preventing local successes from masking global drift.
• Constraint Satisfaction iterates through competing requirements, such as scheduling or resource allocation, until all conditions are met.

How it works

1. Initialize the Loop using the loop-ledger to set a goal, max iterations, and the specific exit mode (manual, target, or dry).
2. Execute and Observe the first step or current state to gather ground-truth data from the environment.
3. Evaluate and Replan by comparing results against the goal; if a conflict or new information appears, the agent generates a revised sequence.
4. Monitor for Stalls using the ledger to detect if the agent is repeating the same failed plan or oscillating between identical constraint violations.

Frameworks & tools

This skill is designed for agents using loop-ledger for iteration tracking and state management. It is framework-agnostic but requires an environment capable of handling serialized state and multi-step execution cycles.

Why this beats prompting it yourself

Standard prompts often result in agents "hallucinating" progress by simply finishing a checklist even if the goal isn't met. This skill enforces a rigorous ledger-based system that distinguishes between task completion and goal attainment, providing specific wiring for handling mid-execution surprises.

Use cases

• Service Migrations where breaking changes are discovered incrementally during the porting process.
• Subsystem Architecture that requires breaking a massive v1 goal into PR-sized executable leaves.
• Resource Scheduling where jobs must be allocated across machines with competing memory and deadline limits.
• Long-running Explorations that need periodic progress scoring to ensure the agent hasn't drifted off-track.

Known limitations

This skill is not for simple linear tasks where surprises are unlikely. It requires the loop-ledger tool to manage iteration state and prevent infinite loops.