Chapter 6: ReAct and deliberate reasoning
The introduction explains why a coherent decision narrative is not enough and presents ReAct as a way to keep conclusions provisional until evidence challenges them.
After this chapter you will be able to
- ReAct hypothesis evidence loop. Describe ReAct as alternating reasoning with information-gathering action, so a conclusion remains provisional until it encounters signals that could challenge it.
- ReAct evidence loop. Explain how ReAct moves from reasoning to action, observation, and revised reasoning, rather than treating a coherent conclusion as final.
- Early evidence-generating actions. Describe why ReAct uses small actions early, and distinguish actions meant to generate corrective evidence from execution after a decision.
- Deliberation as structural pause. Explain how deliberative reasoning interrupts momentum toward a conclusion so assumptions, alternatives, and causal links can be examined before commitment. Describe why this pause matters even before evidence is collected.
- Coherence mistaken for evidence. Explain why a plausible chain from problem to outcome can feel trustworthy while still depending on untested assumptions. Use the AI pilot’s proposed chain from automation to satisfaction as an example.
- Forces reinforcing narrative dominance. Describe how fluency, group dynamics, and accumulated investment make an established explanation seem more credible and harder to revise.
- Partial confirmation and correction. Explain how a favorable metric can validate only one link in a story while other assumptions remain untested, and state how ReAct changes the story’s status.
- Testing before commitment. Explain why ReAct tests a coherent line of reasoning before an organization becomes committed to it.
- Conditions calling for ReAct. Explain why delayed feedback, ambiguous causality, large-scale consequences, and strong social commitment make ReAct valuable. State that these conditions raise the risk of confidence outrunning evidence rather than guaranteeing failure.
- Lightweight hypothesis scaffold. Describe how a team can make its reasoning testable by recording a hypothesis and, in advance, the signals that would confirm or weaken it.
- Advance disconfirming signals. Explain why teams define disconfirming signals before work begins and distinguish such a signal from merely missing a success target.
- Uncertainty reducing test actions. Describe how to choose a test action that challenges the underlying explanation, including when a small experiment or focused data query is enough.
- Preagreed adjustment rule. Explain how an adjustment rule completes the loop and why confidence in an explanation remains provisional until evidence has tested it.
- Provisional automation hypothesis. Explain why the team’s plausible case for automating first responses was not enough to justify rollout, and identify the assumption it needed to test.
- Early disconfirming evidence loop. Describe how the team could agree on disconfirming signals, run a limited test, review results within an observation window, and set a response before expanding automation.
- Timing of corrective feedback. Explain why pulling feedback forward changes the cost of an error in the AI pilot.
- Boundary before commitment. Explain when ReAct calls for a test and why that intervention matters.
- Action as learning. Describe how ReAct uses limited actions to manage uncertainty and build confidence.
What it builds on
- What chain-of-thought actually is (chapter 3) · Assumes the earlier account of making reasoning steps visible.
Section by section
Introduction
The introduction explains why a coherent decision narrative is not enough and presents ReAct as a way to keep conclusions provisional until evidence challenges them.
- Coherent narratives and evidence: Explain why a convincing chain of reasoning can still lead teams astray, especially when assumptions become accepted and early signals are interpreted selectively.
Why it matters: A persuasive explanation should not be mistaken for a tested decision. - ReAct hypothesis evidence loop: Describe ReAct as alternating reasoning with information-gathering action, so a conclusion remains provisional until it encounters signals that could challenge it.
Why it matters: The loop gives a decision a chance to change in response to evidence rather than stopping when its story sounds complete.
Key terms: coherent explanation, assumptions, evidence, ReAct, reasoning and acting, hypothesis, provisional
6.1 What ReAct actually is
This section explains ReAct as a way to test reasoning against reality through small actions and observations before a plausible conclusion becomes a firm commitment.
- ReAct evidence loop: Explain how ReAct moves from reasoning to action, observation, and revised reasoning, rather than treating a coherent conclusion as final.
Why it matters: A convincing explanation can still need revision once it meets evidence. - Early evidence-generating actions: Describe why ReAct uses small actions early, and distinguish actions meant to generate corrective evidence from execution after a decision.
Why it matters: A small test can reveal a faulty assumption before the organization becomes committed to it. - Support rollout assumption failure: Explain what the customer-support AI rollout revealed about a plausible but untested assumption.
Why it matters: Meeting the chosen metric does not prove that the underlying customer need was understood.
Key terms: ReAct, working hypothesis, Act, small loops, prototype, limited release, constrained experiment, evidence, core assumption, response times
6.2 Deliberative reasoning: the discipline of pause
This section explains how a brief pause keeps a plausible explanation provisional long enough to examine its assumptions and causal links before committing to it.
- Deliberation as structural pause: Explain how deliberative reasoning interrupts momentum toward a conclusion so assumptions, alternatives, and causal links can be examined before commitment. Describe why this pause matters even before evidence is collected.
Why it matters: A short pause can prevent a convincing story from becoming a decision before its reasoning has been tested. - Incident narrative and hidden vulnerability: Trace how the post-incident team's initial explanation changed when it reconstructed the timeline, assumptions, and dependencies. Explain the risk of accepting its first coherent story.
Why it matters: The incident shows how an apparently sensible fix can leave the vulnerability that made the problem possible.
Key terms: deliberative reasoning, narrative completion, assumptions, causal links, configuration change, monitoring gap, underlying vulnerability
6.3 Why narrative dominates human decisions
This section explains why coherent organizational stories can feel more convincing than the evidence warrants, and why ReAct treats those stories as hypotheses open to correction.
- Coherence mistaken for evidence: Explain why a plausible chain from problem to outcome can feel trustworthy while still depending on untested assumptions. Use the AI pilot’s proposed chain from automation to satisfaction as an example.
Why it matters: A convincing plan is not the same as a tested one. - Forces reinforcing narrative dominance: Describe how fluency, group dynamics, and accumulated investment make an established explanation seem more credible and harder to revise.
Why it matters: These pressures can make a story harder to question without making it more accurate. - Partial confirmation and correction: Explain how a favorable metric can validate only one link in a story while other assumptions remain untested, and state how ReAct changes the story’s status.
Why it matters: A positive early signal should not close off questions about the rest of the decision.
Key terms: narrative, coherence, untested assumptions, AI pilot, fluency, narrative dominance, social dynamics, political capital, metrics, partial confirmation
6.4 What goes wrong without evidence loops
This section explains how decisions go wrong when evidence cannot challenge a convincing narrative before commitment, and why ReAct tests reasoning while it can still change.
- Testing before commitment: Explain why ReAct tests a coherent line of reasoning before an organization becomes committed to it.
Why it matters: Testing at this point can make correction possible before a decision becomes costly to revise. - Assumptions fading from view: Describe how assumption drift lets a conclusion persist after its supporting conditions stop being questioned.
Why it matters: An established decision may still depend on conditions no one has recently checked. - Narratives selecting evidence: Explain how an established narrative can give supporting evidence more weight than contradictory signals.
Why it matters: A persuasive success metric can obscure other signals that challenge the decision. - Measurement without correction: Distinguish diagnostic use of evidence from performance theatre in reporting.
Why it matters: Having plenty of data is not the same as letting evidence influence a decision. - Costs of delayed recognition: Trace why discovering disconfirming evidence late makes a flawed decision harder to revise.
Why it matters: The cost of a mistaken decision can grow while an organization waits for unmistakable consequences.
Key terms: ReAct loop, testing before commitment, coherence, commitment, assumption drift, assumptions, confirmation selection, narrative, supporting signals, contradictory evidence
6.5 When ReAct is necessary
This section explains when ReAct is especially valuable: when delayed or unreliable feedback lets a convincing explanation gain momentum before it has been tested.
- Conditions calling for ReAct: Explain why delayed feedback, ambiguous causality, large-scale consequences, and strong social commitment make ReAct valuable. State that these conditions raise the risk of confidence outrunning evidence rather than guaranteeing failure.
Why it matters: The point of ReAct here is to keep confidence from getting ahead of evidence. - Delayed and ambiguous feedback: Explain why delayed outcomes call for smaller evidence loops and why ambiguous causality makes apparent improvement an unreliable confirmation of the original reasoning.
Why it matters: Early signals are useful, but their timing and meaning both affect how much confidence they warrant. - Scale and social commitment: Explain how broad deployment raises the cost of untested assumptions and how public commitment can reduce the authority of contrary evidence.
Why it matters: Testing early is especially important before an idea becomes both widespread and difficult to challenge.
Key terms: ReAct, reliable feedback, evidence loops, delayed feedback, ambiguous causality, smaller evidence loops, partial confirmation, scale, untested assumptions, social pressure
6.6 Example patterns and interventions
This section gives teams a lightweight ReAct structure for testing provisional explanations against evidence and changing course when warranted.
- Lightweight hypothesis scaffold: Describe how a team can make its reasoning testable by recording a hypothesis and, in advance, the signals that would confirm or weaken it.
Why it matters: A small written structure gives the team something more reliable than memory or a persuasive story to revisit. - Advance disconfirming signals: Explain why teams define disconfirming signals before work begins and distinguish such a signal from merely missing a success target.
Why it matters: Agreeing on a challenge to the plan beforehand makes it harder to protect a compelling narrative when contrary evidence appears. - Uncertainty reducing test actions: Describe how to choose a test action that challenges the underlying explanation, including when a small experiment or focused data query is enough.
Why it matters: A focused test can challenge a costly assumption without requiring a full rollout first. - Preagreed adjustment rule: Explain how an adjustment rule completes the loop and why confidence in an explanation remains provisional until evidence has tested it.
Why it matters: A preagreed response makes correction part of the decision process rather than an admission made under pressure.
Key terms: hypothesis, confirm, weaken, externalize assumptions, disconfirming signal, circuit-breaker, test action, evidence loop, constrained experiment, focused data query
6.7 Revisiting the AI pilot through a ReAct lens
The section revisits an AI pilot to show how a ReAct evidence loop could test its assumptions before a full rollout makes correction costly.
- Provisional automation hypothesis: Explain why the team’s plausible case for automating first responses was not enough to justify rollout, and identify the assumption it needed to test.
Why it matters: A convincing explanation can conceal the assumption on which a decision depends. - Early disconfirming evidence loop: Describe how the team could agree on disconfirming signals, run a limited test, review results within an observation window, and set a response before expanding automation.
Why it matters: Agreeing on the test and response beforehand lets the team change course without renegotiating direction under pressure. - Timing of corrective feedback: Explain why pulling feedback forward changes the cost of an error in the AI pilot.
Why it matters: Finding a mistaken assumption while a change is small is less disruptive than reversing a scaled decision.
Key terms: initial hypothesis, provisional, key assumption, disconfirming signals, controlled subset, observation window, defined threshold, feedback, loop, scaled
6.8 ReAct as the boundary between reasoning and reality
This section explains how ReAct places an evidence-based test between a convincing explanation and a firm commitment, so organizations can correct course before they commit too far.
- Boundary before commitment: Explain when ReAct calls for a test and why that intervention matters.
Why it matters: A plausible plan can gain authority before evidence has had a chance to correct it. - Action as learning: Describe how ReAct uses limited actions to manage uncertainty and build confidence.
Why it matters: Small learning actions can expose a mistaken direction before it causes larger misalignment. - Culture of evidence-led adjustment: Explain what leaders must make acceptable for evidence to shape decisions.
Why it matters: Tests are more useful when teams can respond to their results without treating revision as defeat. - Reasoning quality before testing: Explain why ReAct still needs clear reasoning and what it adds to earlier reasoning modes.
Why it matters: Testing cannot reliably correct a decision if the hypothesis or the interpretation of evidence is unclear.
Key terms: ReAct, boundary between explanation and commitment, correction, instruments for learning, controlled exposure, hypotheses, confidence, visible incompleteness, learning, individual authority
6.9 Summary
Where this is used later
- From techniques to stack (chapter 7) · Provides ReAct's distinct contribution when combining reasoning techniques into a stack.
- Problem shape as an architectural signal (chapter 7) · The need to connect reasoning with reality.
- When the right model still fails (chapter 7) · Informs why a suitable reasoning model may still depend on the surrounding environment.
Concepts explained in this chapter
- disconfirming signalAn observation that challenges an assumption or hypothesis rather than supporting it.
- narrative dominanceA reasoning failure in which an established decision is defended through a coherent story that filters subsequent information to preserve it.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
