I Ran 157 Agent Plans Against a Real LLM. The Problem Wasn't Execution. It Was Planning.
I thought I was building a better planning engine. What I actually built was a machine for showing me how often a decent-looking plan is still wrong in exactly the way that hurts: not obviously wrong, just missing the one dependency or ordering constraint that turns a migration into an incident. The Failure Starts Before the First Tool Call Your agent can execute perfectly and still fail, because…
I ran an experiment pitting a 157-agent system against a real large language model (LLM). The surprising revelation was not about execution, but planning. My agent could execute tasks flawlessly, yet still fail catastrophically due to flawed planning. The core issue lay in the ordering of dependencies and constraints within the plan.
A goal like "migrate this service to the new auth provider" underwent a chain-of-thought process, only to stumble at step three when the rollback path was missing. This illustrates the danger of optimizing execution layers first, when planning failures often occur beforehand. To address this, I created PlannerCritic, a system that treats a plan like a pull request, with a separate reviewer LLM checking the structure and safety of the drafted plan.
Deterministic gates analyze aspects like ordering, branch sanity, rollback coverage, verification, and preconditions before approving the plan. The critic operates independently from the planner to prevent injection vulnerabilities. The system is designed to fail gracefully by prompting a human if the loop can't converge. The experiment covered 156 scenarios across 35 domains, with 157 execution traces.
While most balanced goals were approved, strict goals were escalated in every case. Adversarial goals also consistently resulted in escalations. This demonstrates that even well-intentioned plans can be fatally flawed, emphasizing the importance of rigorous planning before execution in AI systems.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
