- 1: Part 1, Understanding AI agentsPart 1, Understanding AI agents
- 2: Part 2, The six principles
- 3: Part 3, Introducing an agent
- A–E: Appendices
- 1.1: What an agent is
- 1.2: What an agent works with
- 1.3: How an agent carries out work
- 1.4: What agents are used for
- 1.5: The opening hours example
Part 1, Understanding AI agents
Part 3, Introducing an agent
Chapter 1.3
How an agent carries out work
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.
In this chapter
This chapter explains how an agent works through a task, how its actions are controlled, how several agents can share a task, and how agents go wrong.
How an agent works through a task#
Agents generally work through a task in a loop. The agent interprets the goal, chooses an action, uses a tool where needed and assesses the result before deciding what to do next. Some agents first draw up a plan and then use the same loop to carry out each step, revising the plan where needed. The loop ends when the agent returns the result, asks for help or approval, or reaches a limit set in the harness.
Each step builds on the last. An agent that can check its own results, for example by running tests, can correct some errors as it goes. An early error that it cannot detect may carry into later steps.
Permissions and controls#
Permissions and controls set what an agent is able to do. They are enforced by the systems it uses rather than by the agent itself, so they hold even if it misreads or fails to follow an instruction. They are sometimes called guardrails.
The systems and information that the agent’s account can reach, and whether it can only read them or also change them.
Points at which a person must approve an action before it happens.
Caps set outside the agent, such as a maximum number of steps or a maximum cost for each task.
A log of what the agent did, which tool it used and what changed.
A way for a person to pause or stop the agent and see what it has already done.
Several agents working together#
A task can be divided between several agents. One agent can direct the work and pass parts of it to others, which are sometimes called sub-agents. Another agent can check the work.
Some of these agents are set up in advance for a particular kind of work, each with its own instructions and permissions. Others are created by the directing agent as the task proceeds. It decides what each should do and writes its instructions, so you set the task but not the agents that carry it out.
Each agent should have no more permission than its part of the task requires. An agent created during a task usually works within the permissions of the agent that created it, so those permissions set the limit for all of them.
An error, or a hidden instruction picked up by one agent, can pass to the others and into the combined result, which is usually all you see.
An agent that checks another’s work may share its weaknesses, particularly if both use the same model.
Keep a record of which agent did what, so that a problem can be traced to its source.
Shared standards now allow agents in different organisations to exchange work, but they are at an early stage. Any such exchange would need approval through the organisation’s own process for sharing work or information with another organisation, covering what can be shared, how it is kept secure and who is accountable.
How the harness affects agent performance#
An agent harness is the working structure around AI models. It sets the stages of work, decides which tools can be used, defines what must be saved and checked, and determines when the task is complete.
Stages and sequencing
Tool use
Memory and saved state
Verification and repair
Delegation
Stopping rules
Agent failures are not always failures of intelligence. They can also be failures of structure, such as lost context, weak sequencing, poor checks, unclear roles or no durable record of what has already been done.
The same model can therefore behave quite differently in different products. When you compare products, compare how each handles these parts as well as the model it uses.
How agents go wrong#
Agents can make errors, and an early error may carry into later steps. Knowing the common kinds of error helps you decide where to check.
Stating something false
A model can produce a false claim, source or figure as fluently as a true one.
Giving different answers
The same request can produce different results on different occasions.
Losing track
In a long task, the agent can lose an earlier instruction or decision.
Interpreting instructions literally
An agent can follow the literal wording of an instruction without achieving the result that was intended.
Reporting success too early
An agent can report that a task is complete when part of it has failed or been skipped.
Following hidden instructions
Text in an email, document or web page can contain instructions that the agent follows. This is a security risk known as prompt injection.
© 2026 Considered Ideas Ltd. Licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). This guide provides general information and is not legal or regulatory advice. The guide’s home page gives the suggested credit.
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.

