Specifying the expected behaviour
What the agent must do, what it must never do, when it must stop and ask. Writing the system instruction files and the scope rules.
Align behaviour, do not impersonate a human
An agent should reason like a competent professional: refuse when unsure, escalate instead of inventing, ask for a clarification rather than assume, stay within its scope. At the same time it must be clearly identified as a machine, in line with the transparency obligations of the European regulation on artificial intelligence.
These two requirements do not conflict. The first concerns the quality of judgement, the second the identity of the sender.
What I do
Refusal and escalation rules
The cases where the agent hands back control rather than producing an answer. The most neglected point in production, and the first cause of incidents.
Behaviour evaluation set
Test cases written before going live, the expected behaviour for each, measurement of deviations at every version, behavioural non-regression tests.
Compliance by architecture
Where the notice to the person, the marking of generated content, human validation, the logs and their retention period are placed. These points are decided at design time.
Documentation and traceability
Instruction files versioned in the repository, session logs kept, dated architecture decisions.
Formats
- Scoping — 5 days.
- Prototype — 15 to 25 days.
- Production rollout, on quotation.
Delivered : system instruction files, scope rules, behaviour evaluation set, architecture diagram showing the human validation points and the logs, dated architecture decisions.
The link with the research tax credit
Characterising an expected behaviour, writing an evaluation set, measuring deviations, correcting and starting again is an experimental approach. When the result is not known in advance and the solution cannot be deduced from the state of knowledge, this work may fall under research and development, provided the instruction files, the evaluation sets and the successive versions are dated and kept.
Research and innovation tax credit: software and AI projectsTransparency obligations
Article 50 of Regulation (EU) 2024/1689 covers transparency obligations, in particular informing people who interact with an artificial intelligence system and marking artificially generated content. These obligations have applied since 2 August 2026.
My role is to design the architecture so that these obligations can be met, and to document the choices made. I do not certify the compliance of a system.
An agent to specify, a behaviour to correct
Describe the context and the question. The answer will state what is feasible, in what form, and in how many days.
Get in touch