Frame the decisions
Define the decisions the agent may take, the tools it may call and the actions that always require a human approval.
Build AI agents that act safely in production
An agent that survives production is not a prompt loop. It has explicit tools, permissions, limits, memory, evaluations and a recovery path.
I design and build tool-using agents, including multi-agent orchestration with typed tools in Pydantic AI, LangGraph when the control flow warrants it, and a managed runtime such as AWS AgentCore when operations call for it.
The situations this answers
- An assistant that answers well but cannot act on business systems.
- Agents with no permission model, no limits and no recovery path.
- No evaluation of tool calls, hallucinations or silent failures.
- A single model dependency with no fallback when it degrades.
How it fits together
A supervisor with bounded tools
A supervisor agent calls retrieval, business APIs and a code sandbox, surrounded by memory, an evaluation set and human approval.
What you get
- Agents with explicit tools, permissions and bounded actions.
- Multi-agent orchestration where each agent has one responsibility.
- Evaluation sets for tool use, extraction quality and failure cases.
- Observability on every run, with traces and cost per operation.
- A supervision path and a documented recovery procedure.
Production figures
- Agent workflows in production
- Vector · Soraya 2 public AI assistants in production, designed and operated in-house — plus document AI agents in production at a European energy company
How an engagement runs
Build and evaluate
Implement the graph, the tools, the memory and the evaluation set, then compare architectures on your real cases before scaling.
Operate and hand over
Add monitoring, fallbacks and alerts, run a supervised period, then hand the operating runbook to your team.
Proof and related reading
Frequently asked questions
When is an agent the right choice?
When the task requires choosing tools or adapting a sequence to the content. A deterministic pipeline stays cheaper and more predictable for stable, well-specified processing.
Which frameworks and runtimes do you use?
Pydantic AI for typed agents and tools, LangGraph when the graph needs explicit state, and managed runtimes such as AWS AgentCore when the operating context favours them.
How do you prevent an agent from taking a wrong action?
Actions are allow-listed, arguments are validated, risky steps require confirmation, and every call is traced. Recovery is part of the design, not an afterthought.
How do you measure agent quality?
Each agent gets an evaluation set built from your cases: tool choice, argument correctness, extraction quality and failure handling. Regressions are caught before deployment.
Can agents run inside our cloud?
Yes. Agents can run serverless or on managed services inside your account, with your model providers and your access controls.