Agent Planning Failure Modes

Plans fail via invalid tools, bad parameters, wrong values, unmet goals or constraints, or false success from broken reflection.

Beyond general LLM failure modes, LLM agents fail in planning in distinctive ways.

Tool-use failures (most common):

  • Invalid tool — plan references something outside the tool inventory.
  • Valid tool, invalid parameters — wrong arity or schema.
  • Valid tool, incorrect parameter values — schema-valid but semantically wrong, including argument hallucination of placeholders never stated in conversation.

Goal failure: the plan does not solve the task, or violates constraints (wrong destination, blown budget). Time is an often-overlooked constraint — a late-but-correct grant proposal still fails. In long interactive episodes, Act-without-thought trajectories often fail here by never decomposing the goal into subgoals or by losing track of environment state — the failure mode that sparse ReAct thoughts are meant to cut. Knowledge- intensive ReAct can still fail by repeating prior thoughts/actions without choosing a new next step or exiting — a structural reasoning error that greedy decoding exacerbates; treat loop detection and step-budget fallback (internal–external knowledge routing) as harness guardrails. Trajectory audits also show skipped prerequisites (Act cleaning without navigating to the sink; ReAct-IM believing an object is already clean because the thought said “find a clean …”) and under-commitment after retrieval (Finish NOT ENOUGH INFO when evidence REFUTES). Wrong Thought scaffolds are as dangerous as missing Thoughts.

Reflection-driven failure: the agent wrongly believes it succeeded (assigns 40 of 50 people and insists it is done). Broken reflection stops retries early.

Mitigations are harness-level: clearer tool docs, simpler APIs, stronger planners, plan-validate-execute, and datasets of (task, inventory) pairs that measure valid-plan rate and per-flavor tool-call error rates. Runtime tracing of these failures for production monitoring belongs to observability practice; the design vocabulary above is what you build to reduce them.

Sources

  • AI Engineering (Chip Huyen), ch. 6
  • Prompt Engineering for LLMs (Berryman), ch. 8
  • ReAct (Yao et al.), pp. 1–15