Building and operating AI4 min read

Choose an Agent Pattern

How to choose between assistance, delegated execution and multi-agent workflows, and evaluate the whole task.

A procedure binder guides a bounded agent workstation with explicit tools and permissions.
On this page
  1. 1.Understand the workflow
  2. 2.Think about reliability across steps
  3. 3.Choose a pattern
  4. 4.Set the human role
  5. 5.Keep the system operable

A research assistant can summarise documents supplied by a person. A more autonomous system might search permitted collections, compare conflicting sources and decide when it has enough evidence to draft an answer. The extra steps can reduce work for the user. They also create more places where context, access or execution can fail.

Decide what job the product should perform, then consider how much freedom the system needs to do it.

Understand the workflow

A fixed sequence runs predetermined steps. An agentic workflow lets the system choose among actions, such as searching, calling a tool, asking for information or stopping, based on intermediate results. The distinction is useful when deciding how to test the path. A simpler fixed workflow may be the better product.

A model does not need to change external state to choose a sequence of research actions. Once it can send a message, edit a file or update a record, however, the team must account for the consequences of the action, including partial completion and recovery.

Think about reliability across steps

If five required steps each succeed independently with probability 95%, and every step must succeed without recovery, the chance of all five succeeding is approximately 77%: 0.95^5. This is an illustration, not a measured agent success rate.

Real failures may be correlated, retries may help or make things worse, and some errors can be detected and corrected. Measure the actual workflow with representative cases. Inspect both the result and the path, then decide what quality is needed for the consequence of failure. The evaluation chapter explains how to build that evidence.

Choose a pattern

Three work cells represent research, task execution and manager-worker agent patterns.

Research and synthesis

The system gathers information from permitted sources and produces an artefact for a person to inspect. Decide which sources are suitable, how conflicts are presented and whether the reader can trace claims back to evidence. Read-only access reduces some action risks but does not remove privacy, prompt injection or misinformation risks.

Bounded execution

The system can act through tools: for example, create a draft, run a test or update a permitted record. Define which actions are allowed, whether approval is required, and how the system reports partial completion. Enforce permissions outside the model. Test the failure and recovery path before expanding the boundary.

Multiple agents

A system may divide a task among several workers or run parts in parallel. This can help when work is separable or requires different contexts, but adds coordination, duplicated context, review and cost. Compare it with a single agent or a deterministic workflow on the actual task. Use the additional structure when the measured benefit justifies the operating burden.

Set the human role

A person may choose each action, approve a proposed plan, handle exceptions or monitor a bounded background workflow. Design the interaction for that role: show what is planned or completed, what needs attention and what can be stopped, corrected or undone.

A draft can be low consequence; sending it to a customer may not be. Set the boundary for each action from evidence, reversibility and impact. Some jobs should remain assisted even if more automation becomes technically possible.

Keep the system operable

The agent needs relevant context, tool contracts, permissions, state where required, limits, monitoring and an owner. Long-running work needs a way to resume or hand off without losing what happened. Retrying a failed tool call must account for the possibility that the action already completed.

Estimate the whole trace, including tools, retries, model usage, human review and recovery. The harness chapter covers those components; agent ownership covers the responsibility after release.

v3.1 · Updated Oct 2026