AI agents: a toolbox to stay in control
Asking AI to analyse a website and suggest fixes sounds simple. In practice, you need to define what it can inspect, which tools it uses and where approval is required. A toolbox is the set of tools built around a task, with precise functions and results that can be examined.
AI & automationBy Julien BaillyText updated on 2 min estimated read

Key takeaway
A useful agent relies on understandable tools, a limited scope and success criteria. Approvals are part of the process, particularly when a step changes files, data or a live website.
Define a task with a clear endpoint
An agent can choose its next steps from its tools' results. This requires an observable objective and stopping conditions. “Examine this copy and prepare a report” provides a clearer scope than “take care of the website”.
I start by describing accessible resources, the expected deliverable and situations to escalate: missing access, an incomplete result or a decision outside the scope. These limits also help estimate the time and resources devoted to the trial.
Give each tool an understandable role
Reading a file, comparing references and generating a report are functions that can be described separately. For each, specify the expected inputs, possible effects and result format. An explicit error helps explain what is blocking progress.
Anthropic's work on agent tools emphasises clear descriptions, useful context in responses and evaluations on representative tasks. This design applies both to the toolbox code and the instructions given to AI.
Place approvals where the consequences change
Preparing a plan and modifying an environment have different consequences. A process can begin by analysing a copy, produce a proposal and then request approval before the selected operations. It should be possible to examine the files involved, planned changes and rollback method.
In my WordPress recovery toolbox, still in development, I separate collection, analysis and rebuilding or migration operations. These stages provide a basis for controlled AI assistance; they are not a promise of autonomous recovery to a live state.
Check the process against difficult cases
One working example is not enough to evaluate the whole process. I also prepare cases with missing data, an ambiguous response or a failed command. The expected result may be to stop and ask for clarification.
A useful trace records the tools called, observations, errors and approved actions. It allows the work to be reviewed and the process corrected. These concrete cases determine whether more autonomy brings any value to the user.
Sources and references
Documentation consulted when writing this guide. Advice and procedures should be adapted to your own context.
