Plans rarely survive first contact with reality. The planner generates a multi-step plan based on what it knows, but step 1 might reveal that the file doesn’t exist, the service is down, or the approach won’t work. Before dynamic replanning, the system would just keep executing the remaining steps against assumptions that were already wrong.
Now the planner adapts. After discovery steps — commands that gather information rather than change state — the orchestrator feeds the results back to Claude as structured metadata. The planner can then revise the remaining steps based on what was actually found. A plan that started as “read the config, update the setting, restart the service” might become “the config doesn’t exist, create it from the template, set the value, restart the service” after step 1 reveals the truth.
The non-zero exit code replan extended this further. When a command fails (non-zero exit), the step is marked soft_failed instead of hard-failed. The failure trigger flows through the orchestrator back to the planner, which gets FAILURE RECOVERY guidance in its system prompt. Claude can then generate a fix cycle — diagnose what went wrong, attempt a repair, re-run the original intent.
This was the last feature to land on the mini G-suite benchmark — 12 out of 12 scenarios passing. The benchmark tests real-world task patterns: build a project, debug an error, manage files, configure services. Dynamic replanning meant the system could handle the messy, unpredictable cases that static plans couldn’t.
The system doesn’t just execute plans anymore. It thinks on its feet.