Agent Control Patterns
This is a comparison of control patterns, not a benchmark ranking of products or models. The appropriate pattern depends on task duration, the shape of deterministic work, the cost of unsafe action, and the quality of available feedback.
| Pattern | Strength | Main control | Boundary |
|---|---|---|---|
| Minimal extensible core | Keeps defaults small and exposes local customization | Extensions, session state, user-built tools | Shifts integration and safety work to the operator |
| Stateful tool loop | Supports iterative tool use with caching and compaction | Prompt construction, permissions, context budget | Can accumulate opaque state without observability |
| Long-running planner/generator/evaluator system | Enables multi-session work and role separation | Handoffs, contracts, tests, evaluation | Coordination, cost, and merge pressure grow quickly |
| Programmatic tool orchestration | Keeps intermediate data out of model context | Code-level filtering, concurrency, policy surfaces | Adds orchestration code that must itself be reviewed |
| Permission-gated automation | Reduces prompt fatigue while protecting higher-risk actions | Risk classification and safe fallback | Evaluation error rates and scope remain operational concerns |
Pi illustrates the minimal-core pattern; Codex illustrates a stateful loop; Anthropic’s long-running harness adds evaluator roles; advanced tool use moves filtering into code; auto mode adds permission gates. [src] [src] [src] [src] [src]