AI agents in production.
What it takes to run an agent for real users: knowing whether it works, seeing what it did, limiting what it can do, and what it costs.
Questions
How do you evaluate an AI agent?
Build realistic tasks, run the agent several times on each, grade outcomes with code, models and people, and track capability and regressions separately.
How do you monitor AI agents in production?
Trace every agent run — model calls, tool calls, hand-offs — with inputs, outputs, tokens, latency and errors, then add outcome metrics and human review.
How do you stop an AI agent from taking actions it shouldn't?
Limit an agent's tools, permissions and autonomy, enforce authorisation outside the model, and require a person's approval for high-impact actions.
What drives the cost of running an AI agent?
Agent cost is mostly tokens: how many model calls a task takes, how much context each carries, and which model runs it. How to measure it and bring it down.
How do AI agents manage memory and context?
A model sees only its context window, so agent memory is engineered: saved state, compaction, notes kept outside the context, and clean-context sub-agents.
How should an AI agent recover when a step fails?
Make agent runs resumable: save state after each step, retry only what is safe to repeat, cap retries, and hand the task to a person when limits are reached.
Tell us what you are building.
A short brief gets you a real reply from a founder within one business day — an honest read on fit, scope, and budget.