Amey-Thakur Plan, execute, replan
An agent that plans once and executes blindly follows a stale plan into the ground. One that replans constantly never finishes. The working pattern is an explicit plan, execution against it, and replanning triggered by divergence rather than by every observation.
Method
- Plan in steps with observable outcomes. Each step states what done looks like, which is what makes divergence detectable at all.
- Separate the planner from the executor. The planner holds the goal and the sequence; executors do one step with only what that step needs (see agent-context-isolation).
- Check the outcome after every step. Comparing what happened to what was expected is the entire mechanism; without it you have a script, not a loop.
- Replan on divergence, not on every step. A step that succeeded as expected needs no replanning, and replanning by default burns context and loses coherence.
- Keep the goal fixed while the plan changes. Drifting goals are how long-running agents end up solving a different problem than the one asked (see goal-driven-execution).
- Cap total steps and replans. A hard ceiling turns an infinite loop into a reported failure, which is always the better outcome.
- Surface the plan to the human. A visible plan is reviewable before expensive execution and is the cheapest intervention point (see agent-human-checkpoint).
Boundaries
This pattern handles uncertainty; it cannot rescue an impossible or under-specified goal, which needs a human to reframe. Every replan costs a full planning pass. Steps with irreversible effects must be gated regardless of how confident the plan is (see agent-delegation-protocol).