Plans that change with evidence
AI Agent Planning and Reasoning
Design bounded planning around goals, useful tool calls, revised evidence, stopping rules and observable rationales, with a replacement-part investigation.
AI agent planning is the process of selecting a useful sequence of permitted steps towards a defined goal. Reasoning supports those choices by relating the task to available evidence and constraints. A plan is useful when it improves the next decision and can change when new evidence arrives; a long generated explanation is not proof that the task has been understood or completed.
This guide is for engineering teams designing agents that investigate before acting. The illustrative task is to prepare a documented replacement-part reference for a service team. The agent can inspect approved equipment records, manuals and a substitution register. Qualified staff retain responsibility for technical compatibility and physical work.
Define a goal that has an observable finish
“Find the best replacement” leaves the agent to invent criteria. A stronger assignment is to identify the current approved reference for the verified equipment variant, show the supporting source and flag any conflict or missing approval. The finish can be a supported reference or an explicit unresolved case.
State fixed constraints separately from preferences. The equipment identity and permitted source boundary cannot be traded away to produce an answer. A preference for a concise report can be adjusted without changing the underlying authority. The planner needs to know which conditions are non-negotiable.
Define what the agent cannot conclude. Similar part descriptions do not establish interchangeability. An external catalogue may provide a lead but cannot become an approved substitution merely because it is easier to find. The scoping guide explains how to establish this task boundary.
Choose a planning pattern proportionate to the task
A fixed plan suits a known sequence: resolve the asset, retrieve the applicable manual, inspect the approved substitution record and prepare the result. If every case follows that path, ordinary orchestration may be sufficient even if individual interpretation steps use a model.
A reactive pattern chooses the next step after each observation. If the equipment variant is missing, ask for it. If the manual identifies a discontinued reference, inspect the substitution register. If the source contradicts the asset record, stop for resolution. This avoids committing to a detailed plan before the necessary facts exist.
A plan-and-revise pattern records a short expected route, executes a bounded part and updates it when evidence changes. It can help with longer investigations, provided the plan remains a working artifact rather than a promise that every listed step must be followed.
The ReAct research paper explores interleaving reasoning and actions to incorporate external observations. Its benchmark results are not a guarantee for a business agent; the useful concept here is that observations can change the next step.
Use tools to reduce uncertainty
Before a tool call, identify the information it is expected to provide and how that information affects the task. A search that cannot change the next decision may be unnecessary. This does not require exposing hidden model reasoning; a concise operational purpose is enough for review.
For the replacement reference, an exact equipment lookup resolves identity. A manual lookup establishes the relevant original part. A substitution-register lookup checks whether an authorised replacement exists. Each tool has a distinct evidential role. Calling all available tools simply because they exist wastes time and can widen data exposure.
Use deterministic operations for exact calculations and constraints. A model can choose to inspect a compatibility table, but the application should validate identifiers and preserve units. Where an optimisation or prediction problem exists, use an appropriate method rather than assuming a language model's plan performs the underlying mathematics.
A worked planning record
The following record is a proposed design artifact for the illustrative investigation. It captures externally reviewable state, not a transcript of private reasoning.
| Planning item | Example content | Effect on the next step |
|---|---|---|
| Goal | Prepare the approved reference and supporting evidence | Do not optimise unrelated purchasing choices |
| Verified facts | Equipment identity confirmed in the authorised asset record | Restrict retrieval to that variant |
| Open question | Manual reference may have been superseded | Inspect the approved substitution register |
| Expected observation | Current reference, effective context and approval status | Accept only evidence applicable to the case |
| Conflict | Manual and register disagree about applicability | Pause for source-owner resolution |
| Finish | Supported reference or explicit unresolved report | Do not continue searching merely to avoid an incomplete answer |
Make the record compact enough to stay current. A detailed plan that nobody updates can mislead later steps. Store evidence references and current unknowns separately from narrative summaries so that replanning does not erase why a decision was made.
Replan when the evidence changes the problem
Suppose the initial request names a part, but the asset record shows a different equipment variant. The planner should revisit the identification step rather than continue searching for the originally named part. The goal remains the same; the evidence has changed the route.
Distinguish a failed tool call from a negative finding. “The register is unavailable” does not mean “no approved replacement exists”. “No result for this query” may mean the identifier is wrong or the search method is inadequate. A planner should preserve those distinctions in its next action and final report.
Also distinguish a business stop from a technical retry. A temporarily unavailable source may justify a bounded retry. Conflicting authority may require a person. Repeating the same search cannot resolve an organisational decision that no source has made.
Set stopping rules before the loop starts
Define an execution budget and a completion condition. In an illustrative configuration, the task allows up to 4 evidence lookups and 2 clarification exchanges before returning an unresolved case for review. These are example design limits, not universal recommendations or measured optimal settings.
Track progress as new relevant evidence or resolved uncertainty, not merely the number of steps. Repeating near-identical queries with the same results should trigger a review of the approach or a stop. A loop can appear busy while making no progress towards the finish condition.
The agent should return partial work with clear limitations when it stops. Include the verified identity, sources inspected, unresolved question and appropriate owner. Do not convert a budget limit into a fabricated answer, and do not describe an incomplete investigation as a definitive absence of evidence.
Keep planning separate from execution authority
A model can propose an action that the application must reject. The planner's choice does not override identity, policy or transaction rules. Available tools should be narrow enough that the plan cannot reach unrelated systems or mutate records outside the task.
Retrieved text may contain instructions, but it is evidence to interpret rather than authority to change the assignment. A manual note telling the reader to contact a supplier does not automatically authorise the agent to send a message. The guardrails guide explains how to preserve this boundary.
When a proposal requires approval, show the exact action and supporting evidence. If approval changes the task, update the plan and state deliberately. Do not treat a person's agreement with an explanation as permission for every operation the planner could infer from it.
Use concise rationales instead of unverifiable certainty
A reviewer needs to know why the system selected a source, what evidence supports a conclusion and what remains uncertain. A short decision note can provide that information. It should not claim to reveal the model's complete internal process or present self-reported confidence as calibrated probability.
For example, a useful note says that the proposed reference comes from the applicable approved register and that a conflicting manual version has been escalated. It names observable evidence. A statement that the agent “reasoned deeply” tells the reviewer nothing about correctness.
Keep hypotheses labelled. If the system suspects that an identifier was transcribed incorrectly, it can ask for confirmation or inspect an authorised record. It should not silently rewrite the identifier and proceed as though the hypothesis had been verified.
Evaluate plans through outcomes and constraints
Test whether the chosen sequence obtains the evidence needed for the task, preserves constraints and stops appropriately. Do not require one exact trajectory when several valid routes exist. Conversely, a correct final answer should not excuse an unauthorised retrieval or action along the way.
Include cases with missing identity, stale sources, contradictory documents, unavailable tools and no supported answer. Evaluate unnecessary calls and repeated work as operational quality issues. The testing guide describes how to combine outcome grading with trajectory inspection.
Compare with a simpler fixed workflow. If the adaptive planner mostly follows the same sequence, it may add little value. If it avoids irrelevant work or handles meaningful branching better, measure that improvement alongside latency, resource use and review effort.
Connect the planner to a reliable system
Planning decides what to attempt. Orchestration manages execution state, retries, approvals and recovery. Memory architecture determines which facts and prior context are available. Treating these as separate responsibilities makes the source of failures easier to diagnose.
Multiple planning agents can help when work packages are genuinely separable, but they also require delegation and consolidation controls. The multi-agent guide covers that decision. More planners do not automatically create a better plan.
Open for AI can scope agents for bounded investigation and preparation. To discuss planning for your workflow, bring a case where the next step depends on evidence, the allowed sources and a clear definition of an acceptable unresolved result. Those inputs make adaptive behaviour testable.