Field note
Safe tools can compose into an unsafe workflow
A tool can pass every check on its own and still be dangerous in company.
Agent security and authority.
The claim
A tool can be safe on its own and unsafe in a workflow. Security reviews often examine components one at a time: this model has no training data leakage, this retriever only reads approved documents, this function can only create a draft, this scheduler only runs at night. Each answer can be true. The workflow can still be unacceptable.
The problem is composition. Authority that is harmless in isolation becomes consequential when one component can read, another can write, a third can decide when to act, and a fourth can send the result to someone outside the room. No single tool has all the power. The system does.
The mental model
I find it useful to draw an authority graph rather than a architecture diagram. The nodes are the things that hold data, make judgments, or take actions. The edges are permissions: who can ask whom to do what, with whose identity, under what approval, and with what record afterwards.
Four questions expose most of the risk.
First, what can this set of tools do together that none of them can do alone? Second, whose authority is being delegated when the agent acts, and can that authority be narrowed? Third, where does a human have to make a decision rather than receive a notification? Fourth, what is the audit trail when the workflow crosses three services and two teams?
A tool-level review answers “is this component acceptable?” The workflow review answers “what is this system now capable of, and who is accountable for that capability?” Those are different questions. Only the second one describes the thing you are about to operate.
The evidence
In responsible-AI technical assurance for a global consumer-goods organisation, the useful work was rarely a single-model audit. It was tracing how a model’s output moved through review, deployment and monitoring, and asking which decisions changed because of it. The programme upskilled more than 140 practitioners in fairness, bias mitigation and interpretability, but the assurance outcome that mattered was a changed deployment condition. The review had to affect what went live, not just describe it.
The same pattern appears in regulated public-sector delivery. A pipeline can pass a component test and still fail the operating model when data crosses an environment boundary, when a downstream team assumes an approval that never happened, or when a scheduled job runs with more authority than the task requires. Caldicott, DTN and DPIA obligations are not decoration around the workflow; they are constraints on what the workflow is allowed to become.
Agents make this sharper because they can choose a sequence. A fixed pipeline does what it was programmed to do. An agent with several tools can assemble a path that nobody tested, using permissions that each looked reasonable on their own. That is not a reason to avoid agents. It is a reason to review the composition, not just the parts.
The implication
If you are deploying an agent, threat-model the workflow. Enumerate the actions available to the whole set of tools, not each tool in isolation. Give the agent the least authority that still lets it do the job. Make side effects explicit and reversible where possible. Put human approval at the point where the decision changes, not at the point where the interface looks reassuring. Log the sequence, not just the final answer.
Then test the composition. Ask what happens when a retrieved document contains an instruction, when a tool returns an error, when two approvals arrive in the wrong order, or when the agent decides to retry. Those are normal conditions, not adversarial edge cases.
The question is not “is this tool safe?” It is “what can this set of tools do together, under whose authority, and what stops it?” If the team cannot answer that, the workflow is not ready, even if every component passed its review.