A working agent in production
Orchestration with LangGraph, the OpenAI Agents SDK, CrewAI or plain code, whichever is simplest for the job, deployed behind your auth and your logs.
An agent is easy to demo and hard to trust. I build agents and multi-agent systems that take on a real workflow, call your real systems, and fail safely when the model is wrong, because sometimes it will be.
For product teams adding agents to a product, and operations teams handing a repetitive workflow to software.
01/What you get
Tools wired to your systems
MCP servers and typed tool calls into your APIs, databases, inbox or CRM, with least-privilege credentials for each tool.
Guardrails and fallbacks
Schema-checked outputs, grounding checks, human approval for risky actions and a safe path when the model is unsure.
Evaluation you can rerun
A test set drawn from your real cases and traces in LangFuse or similar, so you can change prompts or models and see what broke.
02/How it works
- Step 1
Map the workflow
We walk through the work as it happens today and mark which steps need judgement and which only need speed.
- Step 2
One narrow agent
I ship the smallest agent that completes one real task end to end, measured against how long it takes a person.
- Step 3
Widen carefully
More tools and more autonomy only where the evaluation numbers say the agent has earned it.
- Step 4
Hand over
Runbooks, dashboards and a session with your team on how to extend the agent and when to distrust it.
03/Proof
Okione, DevSecOps Copilot
Seven agents in a pipeline, with five separate layers that stop hallucinated config reaching a cluster.
Incident Response Agent
Reads customer email, classifies the incident and runs the whole first response in under a minute.
Alex, AI Financial Planner
A planner agent coordinating four specialists on AWS Lambda, SQS and Bedrock.
Orbit, Course Matching & Careers
A live multi-agent advisor that thousands of students use for course and career decisions.
04/Questions
Which agent framework do you use?+
Whichever the problem needs. I have shipped with LangGraph, LangChain, CrewAI, AutoGen, the OpenAI Agents SDK and MCP, and I often end up with less framework than people expect, because plain code is easier to debug.
Which models do you work with?+
OpenAI, Anthropic Claude, Google Gemini, Amazon Bedrock models and open-weight models such as Llama, Qwen and Mistral. The choice comes from cost, latency, data residency and the evaluation results, not habit.
How do you stop an agent doing something harmful?+
Every tool gets the narrowest permissions that work, risky actions wait for a human, outputs are validated before anything acts on them, and every run is traced so a bad decision can be found and fixed.
How long does a first agent take?+
A narrow, production-ready first agent usually takes two to six weeks, depending on how many of your systems it needs to touch.
Related writing