Back to AI Research

AI Research

Adaptive Skill Granularity gives coding agents a route out of failed macros

Key Takeaways

  • Runtime routing chooses between primitive actions and larger skills, with explicit recovery when a macro does not fit.
  • A reusable skill can save an agent several low-level steps, but a large procedure can also fail when its assumptions do not match the task.
  • The [Adaptive Skill Granularity paper](https://arxiv.org/abs/2610.31461) treats the choice between primitive tools and larger skills as a runtime control problem.
  • ## Choose the abstraction during execution
  • The proposed framework conditions skill granularity on task structure, the current execution state and macro-skill failures.

A reusable skill can save an agent several low-level steps, but a large procedure can also fail when its assumptions do not match the task. The Adaptive Skill Granularity paper treats the choice between primitive tools and larger skills as a runtime control problem.

Choose the abstraction during execution

The proposed framework conditions skill granularity on task structure, the current execution state and macro-skill failures. It combines a feature-based controller with a finite-state recovery executor.
This extends the question beyond whether an agent can retrieve a relevant skill before starting. A system also has to decide when the larger procedure remains appropriate and when to return to smaller actions. The useful evaluation target is the transition between those modes, not simply how often a skill is selected.

Compare completion, cost and recovery

The authors use a deterministic coding-agent protocol with verifier-defined success and total cost. They also examine process measures including macro adoption, failure recovery, route match and transition correctness.
The paper reports that fixed macro execution is efficient when applicable but brittle under mismatch. Runtime routing and explicit macro-to-micro recovery improve robustness in the controlled evaluation and recover failures that fixed-granularity execution leaves unresolved. The abstract does not provide a universal productivity multiplier or establish the same outcome in every repository.

Treat a skill as something the agent can leave

For a coding team evaluating this idea, choose a task where a macro’s prerequisites are known. Introduce a controlled mismatch, preserve the execution trace and check whether recovery reaches a verifier-approved result without repeating unsafe or irrelevant steps.
Cost should include the work spent discovering that the macro does not fit, as well as the successful recovery. A shorter trace is not necessarily a better result if it skips validation. The study supports treating skill libraries as controllable and recoverable execution abstractions; it does not turn a stored procedure into a guarantee of correct code.

Comments