XIV · Anti-PatternsAnti-pattern✕

Deliberation Over Observation

also known as Agentic Overthinking, Reasoning-Action Dilemma, Analysis Paralysis, Thinking Instead of Looking

Anti-pattern: give a reasoning-trained model a large thinking budget in an interactive environment and it hypothesises about the environment's state instead of issuing the cheap observation that would settle it.

Context

A reasoning-trained model runs inside a loop that can both think and act, with a real environment attached: a repository and a shell, a browser, a ticketing system, a plant controller. At every step the model chooses freely between extending its internal chain and emitting a tool call, and the thinking budget it is given is large. The post-training that made it good at long chains rewarded self-contained derivations on closed-form problems, where there was nothing external to look at and the whole answer had to come from the chain.

Problem

The model spends that budget reconstructing the environment's state by inference rather than reading it: reasoning about what a file probably contains instead of opening it, about why a test probably fails instead of running it, about what a directory probably holds instead of listing it. Because the chain stays self-consistent it reads like progress, and nothing in the transcript marks the moment an assumption went unchecked. Past a point more thinking lowers task success rather than raising it. Across 4018 SWE-bench Verified trajectories, an overthinking score validated against human expert assessment correlated negatively with resolution rate, and reasoning-tuned models scored higher on that measure than non-reasoning ones.

Forces

  • Internal reasoning is fast, private, rate-limit free and never returns an error, while an environment call is slow, noisy, sometimes rate-limited and can fail outright — so the cheaper-feeling move is the uninformative one.
  • Reasoning-optimised post-training rewards long self-contained chains on closed-form problems, which is the wrong prior for open-world tasks whose ground truth sits one tool call away.
  • A longer chain looks like diligence in a transcript, so the behaviour is rewarded by human reviewers and by trace-length heuristics even as accuracy falls.
  • Cutting the thinking budget is the crude remedy and it also cuts genuine reasoning on the hard cases, so a flat cap trades one failure for the opposite one.
  • Selecting the sample with the lower deliberation-to-action ratio raised measured performance by almost 30% while cutting compute cost by 43%, which makes the quality loss, not just the spend, the reason to bound deliberation.

Example

A coding agent is asked why one test fails. It has a shell and the repository in front of it. Instead of running the test, it spends several thousand thinking tokens reconstructing from memory what the fixture probably sets up, concludes the bug is in a helper it never opened, and edits that helper. The test still fails, and the one command that would have shown the real stack trace was never run.

Diagram

Solution

Therefore:

This entry names the anti-pattern; the corrective is to make observation the default and deliberation the budgeted exception. Bound the thinking a step may spend while a cheap, decisive observation is available, so an assumption about a file, a test result or a directory is resolved by reading rather than by inference. Score each trajectory by the ratio of internal reasoning tokens to environment interactions and treat a high ratio as a defect signal, not as thoroughness; where several samples exist, ship the low-ratio one. Allocate the budget by difficulty rather than flat, so hard cases keep their reasoning while simple ones stop paying for it, and log the step at which the agent last touched the environment so a long chain built on stale observations is visible in review.

What this pattern forbids. No effective constraint is present; the missing one is that the agent must not extend an internal chain past a bounded deliberation budget while an assumption about environment state remains cheaply observable — the observation comes first, and a trajectory whose deliberation-to-action ratio exceeds its threshold is rejected rather than shipped.

The patterns that counter or replace it —

  • alternative-toAdaptive Compute Allocation★— Allocate inference-time compute (thinking tokens, samples, depth, model size) per query based on input difficulty, rather than using a fixed budget across all queries.
  • complementsReAct★★— Interleave a single thought, a single tool call, and a single observation per step so the agent reasons over fresh evidence.
  • complementsPremature Closure✕— The LLM commits to a confident answer before processing all constraints, characteristic of constraint-heavy tasks where it fills in plausible answers fast and gets cross-constraint interactions wrong.
  • complementsOver-Search and Under-Search✕— Anti-pattern: let an agentic RAG system miscalibrate when to retrieve, so it either re-retrieves information already in context or skips retrieval when its parametric knowledge is stale.
  • complementsDecision Paralysis✕— Anti-pattern: when given equally-weighted conflicting goals, the agent either gets stuck trying to satisfy all simultaneously or oscillates between solutions without converging — the most common LLM response to genuine goal conflicts.
  • complementsLarge Reasoning Model (LRM) Paradigm★— Route reasoning-heavy tasks to a reasoning-tuned model that trades inference time for deliberation, rather than to a fast LLM that exhibits premature-closure.
  • complementsToken-Economy Blindness✕— Anti-pattern: operate multi-agent loops with no per-run token budget or alarm, allowing recursive loops to silently accumulate $10k+ in undetected costs.

Neighbourhood

Click any neighbour to follow the language. Scroll to zoom, drag to pan.

References

Provenance