DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents

Who Should Read

Backend developer building LLM agents that finish early with unsupported or guessed answers

Core Mechanics

• One model approving itself is risky
  • In ReAct, the same model decides which actions run and when the task is done.
  • Errors propagate, and agents can stop early with weakly grounded or guessed answers.
• The loop is split into three modules
  • The Brain proposes Thought and Action as in ReAct.
  • The Critic checks syntax, safety and question constraints before execution, and rejects below 0.7 confidence.
  • The Context Manager (CM) rebuilds a STATE from environment-supported facts and outputs SUCCESS only when evidence suffices.
• Weaker models gained the most
  • Qwen3-Coder-480B gained 6.5 to 7.0 Pass@1 points.
  • Claude Sonnet 4.5 gained 4.2 to 5.2 points.
  • Claude Opus 4.5 stayed comparable to ReAct.
• With Opus, grounding improved rather than score
  • On GAIA, evidence completeness and constraint satisfaction rose and admitted guessing fell.
  • Some of ReAct's extra wins came from early completion with weak or guessed answers.

Evidence

• GAIA Pass@1 for weaker models
  • Qwen3 rose from 34.9% (ReAct) to 41.4% (DeReAct), +6.5 points, p=.010.
  • Sonnet gained +4.2 points but p=.170, not significant.
• SWE-bench Verified Pass@1
  • Qwen3 gained +7.0 and Sonnet +5.2 points, both p<.001.
  • Opus scored 65.6%, -0.3 versus ReAct, p=.753.
• Opus grounding on GAIA (vs ReAct)
  • Evidence completeness went from 61.3% to 69.7%, +8.4 points.
  • Constraint satisfaction went from 82.7% to 89.1%, +6.4 points.
  • Admitted guessing went from 21.0% to 15.1%, -5.9 points.
• CM tuning on SWE-bench with Opus Brain
  • A better completion prompt cut FAV POOR from 8% to 2% and moved lift from -4.7 to -1.8.
  • Window W=50 plus an Opus CM gave lift -0.3 with 97% STATE GOOD.

How to Apply

• Add a separate completion-check prompt
  • At the point where your agent emits a final answer, add one extra LLM call.
  • Give it only tool results, not the agent's reasoning, and require evidence before SUCCESS.
• Add a pre-execution Critic gate
  • Reject only dangerous or syntactically broken commands and clear constraint violations.
  • Do not judge strategy, and fail closed when confidence is below 0.7.
  • A smaller model than the Brain is fine; the paper used Haiku 4.5.
• Diagnose gates from trajectory logs
  • Sample about 100 trajectories and check whether STATE keeps evidence and how often unsupported answers are approved.
  • If STATE quality is low, scale the CM model; if unsupported approvals are high, refine the completion prompt.
• Check your history window length
  • Compare the Brain's window to the median ReAct trajectory length.
  • If the window is shorter, widen it.

Try It

# Paste into Claude or ChatGPT (simplified version of the paper's method: a Context Manager completion check as a prompt)

You are a context manager. Judge whether the task is complete using ONLY the command results below. Ignore the agent's own reasoning.
RULES:
- STATUS is SUCCESS only if ALL task objectives are verifiably complete and accurate, with direct evidence.
- Mark data as CONFIRMED only if accessed directly; if you only read ABOUT it, mark INDIRECT.
- Be conservative: when in doubt, use CONTINUE.
- Do NOT suggest next actions.

Task:
{{task description}}

Command results:
{{tool outputs}}

Proposed final answer:
{{agent's final answer}}

Respond in JSON only: {"STATUS": "CONTINUE" or "SUCCESS", "REASON": "...", "STATE": "confirmed facts, constraints satisfied or not, unresolved concerns"}

Next: run it on 5-10 of your agent's answers and check whether any SUCCESS verdict lacks real evidence.

Terminology

Related Papers

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

• Ousted Savannah police officer gave ICE agent husband access to Flock system

  A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired

• Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

  Running Claude Code agents in tmux works well, but limits appear around four agents.