DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents

Ajay Vohra; Tao Chen; Neeti Narayan; Caron Zhang

cs.AIOct 1, 2026

DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents

Authors: Ajay Vohra, Tao Chen, Neeti Narayan, Caron Zhang

Organizations: Amazon · Apple

Abstract

ReAct-based agents typically rely on a single LLM policy to propose actions, interact with the environment, and decide when a task is complete. This coupling makes action authorization and completion control difficult to enforce independently, allowing errors to propagate and unsupported completion claims to terminate execution. We introduce DeReAct, a modular agent architecture that externalizes two gating policies: a Critic that validates proposed actions before execution, and a Context Manager that reconstructs an environment-supported State and certifies task completion. Across GAIA and SWE-bench Verified, DeReAct improves Pass@1 most for weaker Brain models, with gains of 6.5--7.0 points for Qwen3-Coder-480B and 4.2--5.2 points for Claude Sonnet4.5; gains diminish as Brain capability increases. Trajectory and ablation analyses show that external gating is effective when targeted failures are sufficiently prevalent and the gating policy is itself sufficient. With Claude Opus4.5, Pass@1 remains comparable to ReAct, while DeReAct produces more evidence-complete and constraint-satisfying trajectories, indicating that completion control can trade earlier termination for stronger grounding. Overall, DeReAct improves weaker agents while retaining grounding benefits as models strengthen.

Figures & tables

Panel of Figure 1 — Figure 1: DeReAct control flow. The Brain proposes an intent and command; the Critic gates execution; and the Context Manager reconstructs the verified State and determines task completion. Only the Brain additionally receives a bounded history window.

	Pass@1 (%)			Pass@3 (%)
Brain	DeReAct	lift	$p$	DeReAct	lift	$p$
GAIA
Qwen3	$41.4\pm 3.3$	$+6.5$	.010	$55.2\pm 3.9$	$+7.3$	.068
Sonnet	$54.5\pm 3.4$	$+4.2$	.170	$68.5\pm 3.6$	$+7.3$	.060
Opus	$62.6\pm 3.4$	$+1.0$	.697	$73.3\pm 3.4$	$+2.4$	.506
SWE-bench Verified

Table 1: Main DeReAct performance across Brain capability levels. Full results: Table 11 , 12 , 16 , and 18 .

	Critic act. (%)		CM act. (%)		Pass@1 $\Delta$ (pp)
Brain	Abl. B	Full	Abl. A	Full	Abl. B	Abl. A	Full
GAIA
Qwen3	4.3	28.1	92.4	93.3	+0.7	+4.9	+6.5
Sonnet	0.2	2.2	89.2	90.4	+1.2	+4.4	+4.2
Opus	0.1	1.3	54.8	55.7	$-2.2$	+1.0	+1.0
SWE-bench Verified

Table 2: Critic/CM activation rates and Pass@1 lift by module. Full results: Table 9 and 19 .

CM	$W$	Veto	STATE	FAV	Pass@1
		(%)	good (%)	poor (%)	lift (pp)
GAIA
Haiku	25	81.7	80.2	19.8	$-3.0$
Opus	25	38.5	93.9	22.2	$+0.2$
Haiku	50	95.9	79.6	25.8	$-5.0$
Opus ⋆	50	55.7	97.0	16.0	$+1.0$

Table 3: Policy-sufficiency diagnostics for the Opus Brain.

\star

marks the main-results configuration. FAV denotes Final Answer Validation. For SWE-bench Verified, min denotes the one-line base completion criterion and evid the evidence-aware improved prompt ( B.3 ).

Brain	DeReAct	ReAct	lift	R-only	Grounded
GAIA
Qwen3	30.2	23.1	$+7.1$	11.5	55
Sonnet 4.5	43.1	35.6	$+7.5$	23.5	44
Opus 4.5	39.2	35.8	$+3.4$	15.6	20
SWE-bench Verified
Qwen3	43.7	37.3	$+6.4$	17.9	9.2

Table 4: Resolution after CM completion vetoes on matched (instance, attempt) pairs. R-only is ReAct resolution among DeReAct failures; Grounded is the grounded fraction. Full results: Tables 14 and 21 .

Subset	Metric	ReAct	DeReAct	$\Delta$ (pp)
Overall	Evidence complete	61.3%	69.7%	$+8.4$
	Constraints satisfied	82.7%	89.1%	$+6.4$
	Admitted guessing	21.0%	15.1%	$-5.9$
	Both criteria met	60.1%	68.7%	$+8.6$
Correct	Both criteria met	78.4%	82.6%	$+4.2$
Correct	Admitted guessing	5.6%	5.5%	$-0.1$

Table 5: Trajectory grounding on GAIA for Opus Brain. “Both” denotes evidence-complete and constraints-satisfied.

Subset	ReAct	DeReAct	$\Delta$ (pp)
Overall	78.2%	80.8%	$+2.6$
Resolved	80.5%	82.3%	$+1.8$
Unresolved	73.7%	78.0%	$+4.3$

Table 6: Trajectory grounding on SWE-bench for the Opus Brain, measured by appropriate testing. Resolved and unresolved subsets are defined by the benchmark test outcome.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Among correct answers
	Qwen3-480B	Sonnet 4.5	Opus 4.5
Trajectories evaluated	490	476	486
Evidence complete	36.7%	55.7%	61.3%
Constraints satisfied	57.3%	77.7%	82.7%
Admitted guessing	37.6%	23.1%	21.0%
Both criteria met	35.5%	53.6%	60.1%

Appendix

Table 7: GAIA ReAct trajectory grounding across Brain tiers (all three attempts pooled). An independent Opus 4.7 judge, blind to the reference answer, reads each complete trajectory. 2 2 2 Trajectory counts fall below the theoretical maximum (

3\times 165=495

for GAIA;

3\times 489=1{,}467

for SWE-bench) because two exclusion criteria apply. First, runs that exhausted the step budget (

T_{\max}

) or crashed without producing a final answer (GAIA) or a non-empty code patch (SWE-bench) are excluded, since no agent output exists to judge. Second, the small number of trajectories for which the Opus 4.7 judge returned an API error or unparseable JSON are excluded from aggregation. Both criteria filter on structural completeness rather than content quality and are thus plausibly independent of grounding quality; the larger Sonnet shortfall on SWE-bench reflects its higher rate of empty patches (runs completing without producing a diff), consistent with its longer average trajectories hitting the sandbox output timeout.

Among resolved (tests passed)
	Qwen3-480B	Sonnet 4.5	Opus 4.5
Trajectories evaluated	1417	1298	1462
Appropriate testing (overall)	4.6%	80.1%	78.2%
Appropriate testing	8.1%	84.5%	80.5%
Among unresolved (tests failed)
Appropriate testing	2.4%	74.7%	73.7%

Appendix

Table 8: SWE-bench ReAct trajectory grounding across Brain tiers (all three attempts pooled). An independent Opus 4.7 judge, blind to the test outcome, assesses whether the agent performed appropriate testing. † † footnotemark:

			Critic Rej. (%)		CM Rej. (%)
Brain	CM	$W$	DeReAct	Abl.B	DeReAct	Abl.A
Qwen3-480B	Haiku	25	28.07	4.30	93.3	92.4
Sonnet 4.5	Haiku	25	2.20	0.16	90.4	89.2
Opus 4.5	Haiku	25	2.65	0.10	81.7	85.6
Opus 4.5	Opus	25	1.25	0.10	38.5	46.1
Opus 4.5	Haiku	50	4.14	0.10	95.9	95.7

Appendix

Table 9: GAIA Policy Activations. Critic Rej. shows percentage of Brain-proposed actions rejected by the Critic. CM Rej. shows percentage of Brain-proposed final-answers rejected by the CM.

Brain	Config	Level 1	Level 2	Level 3
Qwen3-480B	ReAct	$47.2\pm 6.2$	$33.7\pm 4.3$	$14.1\pm 5.6$
	DeReAct	$48.4\pm 6.2$	$43.0\pm 4.6$	$21.8\pm 6.9$
	Ablation A	$47.8\pm 6.3$	$41.9\pm 4.4$	$16.7\pm 6.2$
	Ablation B	$46.5\pm 6.4$	$34.1\pm 4.3$	$17.9\pm 5.6$
Sonnet 4.5	ReAct	$55.3\pm 6.0$	$55.8\pm 4.9$	$21.8\pm 7.1$
	DeReAct	$62.3\pm 5.9$	$55.8\pm 4.6$	$34.6\pm 8.6$

Appendix

Table 10: GAIA Pass@1 (%) by difficulty level, mean

\pm

standard error over three attempts (

W{=}25

, Haiku CM and Critic).

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$41.4\pm 3.3$	$34.9\pm 3.2$	$39.8\pm 3.1$	$35.6\pm 3.4$	$+6.5$	.010
Sonnet 4.5	Haiku	25	$54.5\pm 3.4$	$50.3\pm 3.5$	$54.7\pm 3.3$	$51.5\pm 3.2$	$+4.2$	.170
Opus 4.5	Haiku	25	$58.6\pm 3.3$	$61.6\pm 3.5$	$58.2\pm 3.2$	$59.4\pm 3.3$	$-3.0$	.201
Opus 4.5	Opus	25	$61.8\pm 3.4$	$61.6\pm 3.5$	$61.2\pm 3.3$	$59.4\pm 3.3$	$+0.2$	1.000
Opus 4.5	Haiku	50	$56.6\pm 3.4$	$61.6\pm 3.5$	$59.6\pm 3.4$	$59.4\pm 3.3$	$-5.0$	.021
Opus 4.5	Opus	50	$62.6\pm 3.4$	$61.6\pm 3.5$	$62.6\pm 3.3$	$59.4\pm 3.3$	$+1.0$	.697

Appendix

Table 11: GAIA Pass@1 accuracy (%), averaged over three attempts. Lifts are relative to the standard ReAct baseline for the corresponding W at the same tier. Values are mean

\pm

standard error.

p

-value from paired permutation test. (GAIA levels breakdown: Table 10 )

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$55.2\pm 3.9$	$47.9\pm 3.9$	$54.5\pm 3.9$	$48.5\pm 3.9$	$+7.3$	.068
Sonnet 4.5	Haiku	25	$68.5\pm 3.6$	$61.2\pm 3.8$	$67.3\pm 3.7$	$63.6\pm 3.7$	$+7.3$	.060
Opus 4.5	Haiku	25	$72.7\pm 3.5$	$70.9\pm 3.5$	$70.3\pm 3.6$	$64.8\pm 3.7$	$+1.8$	.636
Opus 4.5	Opus	25	$72.1\pm 3.5$	$70.9\pm 3.5$	$71.5\pm 3.5$	$64.8\pm 3.7$	$+1.2$	.809
Opus 4.5	Haiku	50	$69.1\pm 3.6$	$70.9\pm 3.5$	$72.1\pm 3.5$	$64.8\pm 3.7$	$-1.8$	.626
Opus 4.5	Opus	50	$73.3\pm 3.4$	$70.9\pm 3.5$	$71.5\pm 3.5$	$64.8\pm 3.7$	$+2.4$	.506

Appendix

Table 12: GAIA Pass@3 accuracy (%), measuring the percentage of questions answered correctly in at least 1 of 3 attempts. Lifts are relative to the standard ReAct baseline for the corresponding W at the same tier. Values are mean

\pm

standard error.

p

-value from paired permutation test.

		Cost ($)				Steps
Brain / CM / $W$	Config	Mean	Median	P90	P95	Median	P95
Qwen3-480B / Haiku / 25	ReAct	0.02	0.01	0.04	0.07	7	20
	DeReAct	3.67	0.35	14.06	17.31	28	1000
	Abl. A	4.32	0.27	16.43	20.45	26	1000
	Abl. B	0.07	0.03	0.16	0.27	8	20
Sonnet 4.5 / Haiku / 25	ReAct	0.57	0.13	1.64	2.48	10	51

Appendix

Table 13: GAIA cost and step distributions per run (USD, list prices, no caching). Means are dominated by a tail of near-limit trajectories; medians characterize typical runs.

			Veto yield				Pass@1 parity
Brain	CM	$W$	Traj. Rej.	Res. $\mid$ Rej.	ReAct all	Yield	ReAct (fail)	Evid. (fail)
DeReAct
Qwen3-480B	Haiku	25	65.5	30.2	23.1	$+7.1$	11.5	55
Sonnet 4.5	Haiku	25	40.8	43.1	35.6	$+7.5$	23.5	44
Opus 4.5	Haiku	25	41.8	51.7	52.7	$-1.0$	23.0	33
Opus 4.5	Opus	25	22.6	39.3	38.4	$+0.9$	14.7	67

Appendix

Table 14: GAIA veto yield. Traj. Rej. : fraction of trajectories with

\geq 1

CM final-answer rejection. Res.

\mid

Rej. : resolution rate among rejection trajectories. ReAct all : ReAct’s resolution rate on the same (instance, attempt) pairs. Yield

=

Res.

\mid

Rej.

-

ReAct all. ReAct (fail) / Evid. (fail) : on pairs the configuration failed, ReAct’s resolution rate and the evidence-complete fraction of those wins (raw counts small; indicative).

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$42.0\pm 0.7$	$37.0\pm 0.7$	$37.6\pm 0.2$	$43.6\pm 0.2$	$+5.0$	${<}.001$
Sonnet 4.5	Haiku	25	$52.7\pm 0.4$	$48.9\pm 0.8$	$53.0\pm 0.0$	$51.3\pm 0.9$	$+3.8$	.008
Opus 4.5	Haiku	25	$61.2\pm 1.0$	$65.9\pm 0.5$	$63.9\pm 1.0$	$66.1\pm 0.2$	$-4.7$	${<}.001$

Appendix

Table 15: SWE-bench Verified Pass@1 resolution (%, averaged over three attempts, base CM prompt ( B.3 )). Lifts are relative to ReAct. Values are mean

\pm

standard error.

p

-value from paired permutation test.

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$44.0\pm 1.8$	$37.0\pm 0.7$	$33.5\pm 0.5$	$43.6\pm 0.2$	$+7.0$	${<}.001$
Sonnet 4.5	Haiku	25	$54.1\pm 0.4$	$48.9\pm 0.8$	$52.2\pm 0.5$	$51.3\pm 0.9$	$+5.2$	${<}.001$
Opus 4.5	Haiku	25	$64.1\pm 0.2$	$65.9\pm 0.5$	$64.3\pm 0.9$	$66.1\pm 0.2$	$-1.8$	.092
Opus 4.5	Opus	25	$63.5\pm 1.0$	$65.9\pm 0.5$	$64.7\pm 1.0$	$66.1\pm 0.2$	$-2.4$	.020
Opus 4.5	Haiku	50	$64.8\pm 0.5$	$65.9\pm 0.5$	$64.6\pm 0.3$	$66.1\pm 0.2$	$-1.1$	.250
Opus 4.5	Opus	50	$65.6\pm 0.8$	$65.9\pm 0.5$	$65.5\pm 0.3$	$66.1\pm 0.2$	$-0.3$	.753

Appendix

Table 16: SWE-bench Verified Pass@1 resolution (%, averaged over three attempts, improved CM prompt ( B.3 ) ). Lifts are relative to ReAct. Values are mean

\pm

standard error.

p

-value from paired permutation test.

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$58.7\pm 2.2$	$54.2\pm 2.3$	$53.8\pm 2.3$	$57.7\pm 2.2$	$+4.5$	.016
Sonnet 4.5	Haiku	25	$63.8\pm 2.2$	$62.8\pm 2.2$	$65.8\pm 2.1$	$64.4\pm 2.2$	$+1.0$	.586
Opus 4.5	Haiku	25	$70.1\pm 2.1$	$70.1\pm 2.1$	$70.8\pm 2.1$	$70.8\pm 2.1$	$+0.0$	1.000

Appendix

Table 17: SWE-bench Verified Pass@3 resolution (%, base CM prompt ( B.3 )). Pass@3 measures the percentage of instances resolved in at least 1 of 3 attempts. Lifts are relative to ReAct. Values are mean

\pm

standard error.

p

-value from paired permutation test.

Brain	CM	W	DeReAct	ReAct	Abl.A	Abl.B	Lift	$p$
Qwen3-480B	Haiku	25	$58.7\pm 2.2$	$54.2\pm 2.3$	$53.8\pm 2.3$	$57.7\pm 2.2$	$+4.5$	.014
Sonnet 4.5	Haiku	25	$66.3\pm 2.1$	$62.8\pm 2.2$	$65.2\pm 2.2$	$64.4\pm 2.2$	$+3.5$	.039
Opus 4.5	Haiku	25	$70.8\pm 2.1$	$70.1\pm 2.1$	$70.1\pm 2.1$	$70.8\pm 2.1$	$+0.7$	.725
Opus 4.5	Opus	25	$69.3\pm 2.1$	$70.1\pm 2.1$	$70.8\pm 2.1$	$70.8\pm 2.1$	$-0.8$	.600
Opus 4.5	Haiku	50	$71.2\pm 2.0$	$70.1\pm 2.1$	$70.1\pm 2.1$	$70.8\pm 2.1$	$+1.1$	.480
Opus 4.5	Opus	50	$71.6\pm 2.0$	$70.1\pm 2.1$	$71.8\pm 2.0$	$70.8\pm 2.1$	$+1.5$	.269

Appendix

Table 18: SWE-bench Verified Pass@3 resolution (%, improved CM prompt ( B.3 )). Pass@3 measures the percentage of instances resolved in at least 1 of 3 attempts. Lifts are relative to ReAct. Values are mean

\pm

standard error.

p

-value from paired permutation test.

			Critic Rej. (%)		CM Rej. (%)
Brain	CM	$W$	DeReAct	Abl.B	DeReAct	Abl.A
Qwen3-480B	Haiku (Base Prompt)	25	22.87	32.55	71.42	77.39
Sonnet 4.5	Haiku (Base Prompt)	25	2.08	1.67	42.91	21.76
Opus 4.5	Haiku (Base Prompt)	25	2.39	2.01	10.18	8.95
Qwen3-480B	Haiku	25	16.15	32.55	93.09	93.08
Sonnet 4.5	Haiku	25	2.06	1.67	72.36	65.29

Appendix

Table 19: SWE-bench Policy Activations. Critic Rej. shows percentage of Brain-proposed actions rejected by the Critic. CM Rej. shows percentage of Brain-proposed final-answers rejected by the CM.

		Cost ($)				Steps
Brain / CM / $W$	Config	Mean	Median	P90	P95	Median	P95
Qwen3-480B / Haiku / 25	ReAct	0.32	0.13	0.55	0.87	34	98
	DeReAct	2.79	1.63	7.16	10.23	154	1000
	Abl. B	1.95	0.50	1.97	3.05	33	93
Sonnet 4.5 / Haiku / 25	ReAct	3.55	2.91	6.84	8.72	52	100
	DeReAct	5.38	2.82	11.66	19.41	68	394

Appendix

Table 20: SWE-bench cost and step distributions per run (USD, list prices, no caching). Mean/median gaps are far smaller than on GAIA because the test harness supplies a natural termination signal; under the Opus Brain with a matched CM, DeReAct reaches mean-cost parity with ReAct.

				Veto yield				Pass@1 parity
Brain	CM	Prompt	$W$	Traj. Rej.	Res. $\mid$ Rej.	ReAct all	Yield	ReAct (fail)	Evid. (fail)
DeReAct
Qwen3-480B	Haiku	base	25	72.3	41.1	35.3	$+5.8$	14.9	5.4
Qwen3-480B	Haiku	impr.	25	96.3	43.7	37.3	$+6.4$	17.9	9.2
Sonnet 4.5	Haiku	base	25	3.4	32.0	26.0	$+6.0$	14.7	20.0
Sonnet 4.5	Haiku	impr.	25	30.3	48.9	39.2	$+9.7$	17.6	70.0

Appendix

Table 21: SWE-bench veto yield. Columns as in Table 14 , with evidence assessed as appropriate testing.

			State		FAV
Bench	Config	Brain / CM / Prompt / $W$	GOOD	POOR	GOOD	POOR
GAIA	DeReAct	Qwen / Haiku / impr. / 25	69.7%	0.0%	36.0%	31.5%
		Sonnet / Haiku / impr. / 25	80.7%	1.1%	53.4%	17.0%
		Opus / Haiku / impr. / 25	80.2%	1.1%	48.4%	19.8%
		Opus / Opus / impr. / 25	93.9%	1.0%	41.4%	22.2%
		Opus / Haiku / impr. / 50	79.6%	1.1%	44.1%	25.8%

Appendix

Table 22: Context Manager decision quality, rated by an independent Opus 4.7 judge reading each complete trajectory (100 sampled per config). State synthesis improves with CM model scale on both benchmarks; Final Answer validation (FAV) does not track scale and instead tracks the domain and prompt.

AI Agent Reliability

AI agent reliability studies whether tool-using systems act correctly, consistently, safely, and recoverably across changing tasks and environments.

See all 339 papers

ReDAct: Uncertainty-Aware Deferral for LLM Agents
Apr 8, 2026Dzianis Piatrashyn, Nikita Kotelevskii, Kirill Grishchenkov +7LLM Agents LLM Agent Reliability
Mohamed bin Zayed University of Artificial Intelligence · Ocapital · National University of Science and Technology (NUST) MISIS +3
Credo: Declarative Control of LLM Pipelines via Beliefs and Policies
Apr 15, 2026Duo Lu, Andrew Crotty, Uğur ÇetintemelLLM Agent Orchestration LLM Agents
Brown University · Northwestern University
Think Before You Act: Decision Gate for LLM Agents
Oct 28, 2026Iftach Shoham, Yuval David, Offer Akrabi +5LLM Agent Verification Runtime Enforcement for AI Agents
IBM SIL

Apr 8, 2026J/K move · Enter open · S save

ReDAct: Uncertainty-Aware Deferral for LLM Agents

Dzianis Piatrashyn, Nikita Kotelevskii, Kirill Grishchenkov, Nikita Glazkov+6

Mohamed bin Zayed University of Artificial Intelligence · Ocapital · National University of Science and Technology (NUST) MISIS+3

LLM Agents LLM Agent Reliability LLM Routing LLM Inference Efficiency

Open paper AskPDF arXiv

Recently, LLM-based agents have become increasingly popular across many applications, including complex sequential decision-making problems. However, they inherit the tendency of LLMs to hallucinate, leading to incorrect decisions. In sequential settings, even a single mistake can irreversibly degrade the trajectory, making hallucinations an even bigger problem. Although larger LLMs hallucinate less, they incur a significantly higher per-token cost. In this paper, we address this tradeoff by proposing ReDAct (Reason-Defer-Act). In ReDAct, an agent is equipped with two LLMs: a small, cheap model used by default, and a large, more reliable but expensive model. When the predictive uncertainty of the small model exceeds a calibrated threshold, the decision is deferred to the large model. We evaluate our approach in text-based embodied environments such as ALFWorld and MiniGrid and show that deferring only about 15% of decisions to the large model can match the quality of using it exclusively, while significantly reducing inference costs.

Abstract

AI Agent Reliability

Explore similar work

ReDAct: Uncertainty-Aware Deferral for LLM Agents

Credo: Declarative Control of LLM Pipelines via Beliefs and Policies

Think Before You Act: Decision Gate for LLM Agents