ReWOO is a planning pattern for AI agents that separates reasoning from tool use. The model first lays out a plan, then executes tool calls with the planned steps, which can reduce repeated model invocations and make tool use easier to audit and control.
How ReWOO Changes Agent Control Flow
ReWOO changes the usual agent loop by separating planning from execution. That means the model can decide the work order first, then carry out tool calls against a prebuilt sequence instead of constantly interleaving thought and action. The design is useful when teams want clearer control flow, fewer repeated calls, and a cleaner place to inspect what the agent intended to do.
Compared with reactive loops, the planning step acts like a coordination layer. It can reduce unnecessary back-and-forth with tools, but it also shifts more importance onto the quality of the initial plan. If the plan is wrong, incomplete, or based on faulty assumptions, the downstream execution will faithfully follow those errors.
Why ReWOO Matters for Tool Use and Auditing
ReWOO is not just a performance pattern. It also changes how tool use is governed because the plan becomes a visible object that can be reviewed, constrained, or logged before execution. That makes it easier to understand why a tool was called, what dependency came first, and where a step may have been unnecessary or unsafe.
This is especially relevant in agent systems that interact with external services, data stores, or privileged operations. A structured plan can improve auditability, but it does not remove the need to validate tool arguments, output trust boundaries, and the permissions behind each action. The plan may be easier to read than a reactive trace, yet it is still only as trustworthy as the inputs and objectives that produced it.
Plan-Then-Execute vs ReAct Agents compares this loop shape with ReAct-style behaviour and shows where control-flow integrity concerns appear in practice.
Common Failure Modes in Plan-Then-Execute Agents
The main failure modes come from treating the plan as if it were automatically safe. A model can generate a plausible sequence that still contains a bad dependency, a misleading assumption, or an unsafe tool request. If later steps depend on earlier outputs, one bad step can cascade through the rest of the run.
There is also a control boundary issue. Because planning and execution are separated, defenders may assume the planning phase is benign and focus only on tool output. In reality, the plan itself can carry risky instructions, hidden assumptions, or prompt-injection influence that shapes what happens next.
For that reason, ReWOO should be understood as a control-flow pattern, not a security control by itself. It can improve observability and reduce repetition, but it does not guarantee safe autonomy, safe delegation, or correct tool selection.
Where ReWOO Fits in Agent Security Design
ReWOO is best used when the agent needs disciplined sequencing, explicit dependencies, and a way to separate reasoning from execution evidence. It is especially valuable when a team wants to inspect or constrain the planned actions before tools are used, or when repeated model calls would otherwise obscure what the agent is doing.
That said, the safer design is usually the one that combines a clear plan with downstream checks on tool scope, argument validation, and output handling. ReWOO can make those checks easier to place, but it does not replace them. In practice, the pattern is strongest when teams treat the plan as a reviewable artifact and the execution layer as a separately enforced trust boundary.
Plan-Then-Execute vs ReAct Agents is the most direct reference for understanding how this architecture changes agent risk, replanning behaviour, and where validation should sit.
Risk and Threat Considerations
ReWOO can reduce repetition, but it also concentrates more trust in the initial plan and the assumptions embedded in it. If an attacker can influence the plan, the agent may carry a bad instruction set through multiple later tool calls, which makes the error more efficient and more persistent than a one-step mistake.
Failure mechanism: Plan content can be poisoned, misdirected, or built on unsafe assumptions, and the execution phase may then amplify the flaw across several coordinated tool actions.
Impact: The result can be incorrect outputs, unsafe side effects in connected systems, or harder-to-detect abuse because the harmful intent is spread across a seemingly orderly sequence.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and MITRE ATT&CK address the attack and risk surface, while NIST CSF 2.0 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI02 — Tool Misuse | ReWOO changes how agents select and execute tools. |
| ASI03 — Identity & Privilege Abuse | Planned execution can carry excessive or misused authority into tools. | |
| Recommendation — Constrain tool calls to approved actions and validate each planned step before execution. Limit agent authority and map each planned action to the minimum required privilege. | ||
| MITRE ATT&CK | T1059 — Command and Scripting Interpreter | Agent tool execution often reaches command-like action surfaces and scripted automation paths. |
| Recommendation — Inspect planned automation for command execution opportunities and monitor for abnormal scripting behaviour. | ||
| NIST CSF 2.0 | PR.AA-05 — Least Privilege | Agent tool access should be limited to the minimum needed for the planned task. |
| DE.CM-08 — Vulnerability exploitation monitoring | Plan-and-execute workflows benefit from monitoring for suspicious tool-use patterns and abuse. | |
| Recommendation — Apply least privilege to the agent’s tool and data access paths. Monitor agent tool activity for unexpected sequencing, reuse, or abuse patterns. | ||
Practitioner Guidance
What to watch for: The key judgement is not whether the agent can produce a plan, but whether the plan is being treated as a governance object. Teams should decide where plan review, tool validation, and execution approval actually happen, especially when the agent can reach external systems or sensitive workflows.
Practitioner note: ReWOO is most useful when it improves traceability without creating false confidence. A readable plan is helpful, but the real control is still the combination of constrained tool access, argument checking, and post-execution review.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 16, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org