Agent contract
Job, allowed inputs, tools, actions, exclusions, and escalation.
Production AI agents · Europe
AI agents that can use tools without hiding their decisions.
I build focused agents for research, knowledge retrieval, support, reporting, and operational follow-up. The system is designed around approved context, bounded actions, evaluation, logs, and a person who remains accountable.
Direct answer
What makes an AI agent production-ready?
A production AI agent has a defined job, controlled access to data and tools, representative evaluations, observable actions, explicit approval points, and a recovery path. The model is one component. Reliability comes from the surrounding software: retrieval, permissions, state, deterministic checks, logging, and an interface that shows uncertainty instead of concealing it.
When this helps
Use an agent when the work requires judgment across changing inputs and bounded tools.
- Research must combine several approved sources and preserve citations.
- A support or operations team needs a draft and suggested action, not an unreviewed decision.
- Recurring reports require retrieval, calculation, synthesis, and publication steps.
- A knowledge system must answer with permissions and source context.
- A multi-step workflow needs state, retries, escalation, and an audit trail.
What the work covers
The product is the control system around the model.
The build covers the complete path from source to action and from failure back to a responsible person.
Context and tools
Retrieval, APIs, credentials, permissions, and state.
Evaluation
Representative cases, quality criteria, red-team cases, and regression tests.
Operations
Logs, traces, cost limits, approval, fallback, and documentation.
Compare the options
Not every automation needs an agent.
| Pattern | Best when | Risk |
|---|---|---|
| Deterministic automation | Rules and inputs are stable. | Poor fit for ambiguous inputs. |
| Single AI step | One classification, extraction, or draft can be reviewed. | Context and quality still need evaluation. |
| Agent workflow | The task requires tools, state, branching, and bounded judgment. | More autonomy increases testing and operating cost. |
| Human service | The decision is consequential, rare, or depends on tacit trust. | May be slower, but accountability remains clear. |
How the work runs
Define the job before choosing the model.
Bound the task
Write the successful outcome, prohibited actions, and escalation rules.
Build the context
Connect approved sources and tools with least-privilege access.
Evaluate
Run representative and adversarial cases before granting actions.
Release gradually
Observe real use, review failures, and widen the boundary only with evidence.
Relevant work
Agentic Brief is a working example.
The system combines reader preferences, LinkedIn enrichment, research, review, and publication. It treats the agent as part of a product and operating workflow rather than a chat window.
FAQ
Questions before building an AI agent.
Which model do you use?
Model choice follows the task, data, latency, tool use, evaluation results, and operating budget. The system should allow a model to be replaced without rebuilding everything.
Can the agent access our private data?
Yes, through explicit sources and permissions. Access should be least-privilege, logged, revocable, and separated by user or organisation where required.
How do you prevent hallucinations?
No single technique prevents every error. Grounding, citations, structured outputs, deterministic checks, abstention, evaluation, and human review reduce the risk.
Can the agent take actions?
Yes, but action permissions should begin narrow. Consequential actions normally require confirmation, limits, and an audit trail.
How much does an agent cost to operate?
Cost depends on model calls, context size, retrieval, tool use, volume, retries, and monitoring. I estimate the operating envelope before recommending production use.
A bounded job, not an AI employee
Bring the task, tools, and decisions the agent would need to touch.
I will help define whether the right answer is an agent, a smaller AI step, deterministic automation, or no build at all.
Discuss the agent