When plans break
Wrong decompositions, stale assumptions, and compounding drift. Why long autonomous plans are unreliable, and how checkpoints turn a broken plan into a recoverable one.
The plan was a guess. Sometimes it is a good guess and the work flows; sometimes the first step returns something that invalidates the whole shape of it. How an agent handles that divergence — re-plan, or grind on — is the difference between a tool you trust and one you watch.
Start here
A plan is a hypothesis about a task you have not done yet. The skill is not writing a perfect plan; it is noticing when reality has diverged from it and re-planning instead of executing a script that no longer fits.
A 8-step chain — click a step to verify it
with no verifiers, 8 steps at 90% each gives 43%.
Compounding is the agent version of a game of telephone: a wrong intermediate result is read as fact by every later step. A verifier at the critical step resets the odds — a tool that checks the arithmetic, a test that fails, a schema that rejects the call. The longer the chain, the more a single checkpoint buys. Numbers are illustrative.
Click steps to add verifiers and watch the compounding reset. Long chains are where small errors become certain failure.
Failure modes
- Wrong decomposition — the subtasks do not actually compose into the goal, so each can succeed while the whole fails.
- Stale assumptions — step 1 discovers the data is shaped differently, but steps 2–8 still assume the old shape.
- Premature commitment — the agent has decided on an approach and reads every new result as support for it.
- Compounding drift — a wrong intermediate result is treated as fact by every later step, so the error is never corrected, only amplified.
Recovery
- Re-plan triggers — define what counts as "reality diverged" (a tool error, a failed check, a value out of range) and re-plan on it rather than pressing on.
- Checkpoints — verify at the steps where being wrong is expensive; each verifier resets the odds of everything downstream.
- Fall back to a workflow — when the open-ended plan keeps failing, the bounded path is the reliable one (the foundations chapter's lesson, now with teeth).
- Escalate — a human is the ultimate re-planner; route the stuck case to one instead of looping.
Careful
Long autonomous plans are the least reliable thing in this domain, and the honest thing is to say so. The published wins are on tasks with cheap, automatic verification — where the agent can tell it is wrong. On open-ended tasks without a checker, the plan drifts and the confidence does not. Keep plans short, verify where you can, and put a human at the checkpoints that matter.
Check yourself
Eduspheria wiki · Agentic AI, Planning & reasoning
0 / 4 answered
That is planning: decompose, reason, reflect, search, and re-plan. Next, the memory that makes any of it survive a long task.