AI agent consulting
Limited advisory availability for teams running coding agents in production.
I work full time and take a small amount of advisory work, by request, on running coding-agent fleets in production. Availability is limited, and I only take questions that fit my own published work.
Teams look for this as AI agent consulting, agentic AI consulting, or an LLM or AI/ML consultant. My part of that is operational: how work reaches coding agents, how it is checked, what it costs, and where a human stays in the loop.
Where I can weigh in
Fleet dispatch. I have worked through the practical problems of assigning tasks to concurrent agents, starting workers without overwhelming a host, and stopping repeat failures. The fleet failure map, worker state machine, and workflow manual show the system and the failure modes behind this area.
Verification gates. An agent's report of success needs an independent check, and a gate that cannot run should not look like failed work. I have written about independent verification, mute gates, and system-level trust from operating those checks.
Cost per closed task. Token price alone does not say whether a fleet is productive. I track completed work, retries, and the cost of the whole run. The unit-economics note, benchmark analysis, and inference plan data give the method and its inputs.
Infrastructure and change boundaries. Agent-run CI needs a clear line between work an agent may start and changes that require a human or a declarative deployment path. The workflow manual, CI on Rackspace Spot guide, and cluster and DevPod guide document first-hand examples.
Fit and contact
I don't build a product for another team, fill a staff role, write paid content, evaluate vendors for a fee, or take work that conflicts with my full-time employment. A useful fit is a team already running coding agents with a bounded operational question that can be answered from public, first-hand experience. If a published note already answers it, I'll point there.
Email [email protected]. A short note can say (1) what you are running, (2) the specific question, and (3) any constraints that change the answer. I read messages, but cannot answer every one. Please leave confidential details out of the first email.