Secure agents and APIs
How can an agent be allowed to act on a business API without delegating critical controls to it?
3500 € per session, excl. VAT
In short
In two days, totalling 14 hours excluding breaks, you will design an agent prototype connected to business APIs in an isolated environment. You will define call contracts, secure authorisations, handle errors and evaluate the actions performed. The programme culminates in a tested demonstration and a pilot plan.
Why this training?
An agent may select the wrong endpoint, invent an identifier or repeat a write operation after a timeout. A plausible response therefore does not prove that the business operation is correct. Separating reasoning, execution and authorisation makes it possible to control the actions actually performed. Strict contracts, sanitised traces and reproducible tests make the prototype’s limitations visible before a pilot. Running case study: The fictitious organisation « Falnorie Distribution » wants to assist its orders department: checking stock, finding an order and proposing a reservation. Every write operation requires human confirmation. The exercises use fictitious data and simulated APIs. Transfer to the workplace: Adapt the prototype to an authorised API: have the permissions validated, rerun the tests in an isolated environment, then submit the pilot plan to the business and security leads.
Objectives of this training
- Define the scope of actions available to the agent.
- Implement call contracts validated on the execution side.
- Apply least-privilege permissions and confirmation of write operations.
- Test errors, injections and repeated calls.
- Measure operational quality and formalise a pilot.
Who is it for and what are the prerequisites?
Roles and tasks concerned
- Developers and integration engineers responsible for business connectors.
- Architects and technical product leads defining the scope of an agent pilot.
Prior skills
- Be able to program, call an HTTP API and work with JSON.
- Understand authentication, HTTP codes and automated tests.
- Have experimented with instructions and structured outputs from generative AI.
What you will work on
Defining the scope of delegation
Objective: Choose the permitted operations. Reading, writing, business invariants and trust boundaries.
Defining contracts for calls
Objective: Eliminate ambiguous arguments. JSON schemas, mandatory fields, errors and versioning.
Building the connectors
Objective: Link decisions to the simulated APIs. Adapters, serialisation, response normalisation and externally stored secrets.
Orchestrating the agent
Objective: Bound the agent’s autonomy. Action selection, clarification, iteration limit and stopping.
Protecting write operations
Objective: Control every reservation. Least-privilege permissions, server-side checks and confirmation tied to the parameters.
Withstanding failures
Objective: Avoid duplicate operations. Timeouts, bounded retries, idempotency and uncertain outcomes.
Testing trust boundaries
Objective: Prevent action hijacking. Injection in API responses, filtering and data minimisation.
Instrumenting execution
Objective: Make actions auditable. Correlation identifiers, sanitised traces, latency and consumption.
Evaluating the prototype
Objective: Measure business outcomes. Normal cases, expected refusals, repetitions and acceptance thresholds.
Preparing the pilot
Objective: Decide on the conditions for experimentation. Supervision, limited scope, shutdown and return to manual processing.
What you take away
- Matrix of actions, permissions and human validations.
- Executable prototype with connectors and launch instructions.
- Test suite and measurement report.
- Pilot plan with stopping conditions and responsibilities.
Method and assessment
- Short instructional sessions followed immediately by practical work on the running case study.
- Individual work followed by work in pairs, with peer review of the outputs.
- Guided group discussion: what works, what needs correcting and what remains a human decision.
- Preparation: Prepare a development environment and AI tools authorised by the organisation, without production secrets. Bring an anonymised API contract or use the fictitious contract provided; check access to the sandbox.
Observable criteria: Every invalid argument is rejected before the business function call. ; No tested write operation bypasses authorisation and confirmation. ; A simulated repeat does not create an additional booking. ; Traces link the request, call and result without exposing any secrets. ; The report distinguishes business success, errors, latency and observed cost..
Frequently asked questions
Is a real API essential?
No, simulated APIs also allow failures to be tested.
Which tools should be used?
Only those authorised by the organisation, in line with its technical constraints.
Does the agent receive the secrets?
No, the connectors use them outside the context passed to the model.
Will the prototype be deployed to production?
No, deploying it to production requires further validation and additional work.
Last updated: