DeReAct introduces a new architecture to enhance reliability in AI agents

Summary

DeReAct splits the monolithic ReAct loop into a Brain and two external oversight policies: a Critic that vetoes bad actions and a Context Manager that confirms task completion. This architecture yields a 4–7 point Pass@1 improvement on GAIA/SWE-bench for weaker models, successfully matching top-tier model scores and improving trajectory grounding.

Why it matters

In production agent deployments, a single large language model (LLM) both proposes and validates actions. In this setup, a confident hallucination can terminate a task with an unsupported 'completed' claim. By separating authorization and completion into distinct, auditable policies, DeReAct offers a practical template for security-critical workflows requiring independent checks, without needing to retrain the core model.

Highlights

• Critic + Context Manager oversight boosts Pass@1 scores by 6.5–7.0 points for Qwen3-Coder-480B on GAIA/SWE-bench Verified.
• The stronger the Brain model, the smaller the gains; with Claude Opus 4.5, Pass@1 scores remain at the same level as standard ReAct.
• Ablations show the architecture is only beneficial if the targeted error mode is frequent enough and the oversight policy is competent on its own.

Critical take

The paper’s own ablations admit the benefit is conditional: oversight only works when errors are 'sufficiently common' and the oversight policy is 'sufficiently competent' on its own. This is a near-tautological success criterion that offers little guidance on where to deploy it. For teams already using frontier-class Brain models, the practical Pass@1 difference is nearly zero. What remains is merely a claim of 'more evidence-complete trajectories' without a clear operational metric for benchmarking.