Agentic AI Engineering
Agentic AI that closes the loop, not just chat.
I design and ship autonomous agents that plan, call tools, recover from failure, and produce verified outputs. LangGraph, Pydantic AI, and hardened tool use.
(01)What you get
Most AI features in production today are chat wrappers. Agentic AI is different: the system decides what to do, uses real tools, validates its own outputs, and retries when it fails. I build these systems end-to-end, from the agent loop and the tool manifest to observability, eval harnesses, and the cost guardrails that keep them shippable.
- 01
Loop + tool design
Concrete graphs with typed state, deterministic tool contracts, and failure paths. No soup of prompts calling prompts.
- 02
Evaluations that matter
Task-specific eval sets, ground truth, and regression harnesses so you can ship changes without praying.
- 03
Cost + latency guardrails
Per-turn budgets, model routing, prompt caching, and fallbacks. Agents that don't bankrupt you when traffic grows.
- 04
Human-in-the-loop where it counts
Clear handoff moments, audit trails, and escalation paths. Autonomy is a dial, not a switch.
(02)How we work
- 01
Discover
Map the workflow, identify the autonomy boundary, and pick the agent topology (single, supervisor, or hierarchical).
- 02
Prototype
Ship a narrow-scope agent in a week. Wire it to real tools, real data, and a real eval set from day one.
- 03
Harden
Add retries, timeouts, cost guards, tracing, and drift detection. Replace magic with observability.
- 04
Operate
Weekly eval runs, incident playbook, and versioned prompts. Changes ship through CI with regression gates.
(03)Stack and engagement
LangGraph, Pydantic AI, OpenAI, Anthropic Claude, FastAPI, Postgres + pgvector, Langfuse, Docker
Engagements run from 2-week pilots to 3-month builds. Day rate or fixed-scope, billed milestones. Exact pricing depends on scope and access to data, let's talk.
(04)Frequently asked
Is Agentic AI ready for production today?
For narrow, well-scoped tasks with good evals: yes. For open-ended 'do anything' agents: not without tight human review. I help you pick a scope that will actually ship and drive business value within weeks.
Which agent framework do you use?
LangGraph for anything with branching, retries, or multi-step tool use. Pydantic AI for typed, single-turn agents. I avoid frameworks that obscure state. Observability is non-negotiable.
How do you control cost?
Per-turn token budgets, model routing (cheap model for classification, strong model for reasoning), prompt caching (Anthropic + OpenAI where applicable), and a kill switch tied to spend anomalies.
Do you integrate with existing systems?
Yes. I treat tools as first-class contracts. Your CRM, database, internal APIs, n8n workflows and webhooks become typed tools the agent can call with validated inputs and schema'd outputs.
What's the fastest we can start?
I typically run a 2-week pilot with a single, well-defined agent. By the end you have a working loop, an eval set, and a clear decision on whether to scale.