π I'm a Forward Deployed Engineer. I embed with teams, learn their domain, and ship agentic systems that survive contact with production. Agent harness design, evals and security are the parts I enjoy most (happy to connect on LinkedIn also)
β‘ Real facts:
"Almost every agent demo works ... the question is what happens on the 200th run, with real data, real permissions and a rate limit in the way." π¬
"The model is the cheapest part of the system. The harness around it is where the engineering actually lives ... tool loops, context budgets, retries, permission boundaries." π§
"An agent without an eval harness is a demo with good luck. You will find out in production, and the user will find out first." π§ͺ
"Silent degradation is the expensive failure mode. A run that half worked and said nothing costs more than one that stopped and explained itself." π
"A workflow that runs is not a service. A service has retries, idempotency, alerting and somebody who owns the pager at 2am." βοΈ
"Multi-agent is not a flex. Most 'crews' are one well scoped prompt with extra latency and three new failure modes ... sometimes the right answer is a cron job." π€
"Using one framework for every agent problem is lazy. Think about the differences, or you will ship a graph where a function would have done." π¨
"Our AI agent is fully autonomous and production ready! Here are the logos of our enterprise customers ..." Oki, challenge accepted. Show me the eval suite, the failure modes and the rollback plan. π
"Everyone puts the API key in .env and calls it security. The real questions are credential scope, blast radius and whether anyone can reconstruct what the agent touched." π
"Trust but verify applies 100% to model output. It is fluent, confident, well formatted and occasionally completely invented."
"Ask an AI whether your startup idea is good and it will find a way to say yes. I built the thing that scores it instead, and one fatal dimension sinks the whole number." π
"Your competitive estimate before research is not an estimate, it is an upper bound. I measured it five times on real ideas ... it fell 20 to 40 points every single time." π
"You cannot solve a domain you have not sat inside. Two weeks with the people doing the work beats two months of requirements documents." πͺ
"Almost everything is possible, it is only time, budget and quality ... but resources are finite, so the actual skill is knowing what not to build." β³
"Everyone ships a broken agent once. Shipping the same broken agent twice means nobody read the logs ... feedback and root cause beat heroics." π΄
"Nobody appreciates documentation until the person who wrote the system leaves." π
"If you cannot explain the failure mode to the client's security team, you do not understand it yet." π£οΈ
An AI venture analyst that runs in your terminal. It researches a startup idea with real sources, scores it out of 100 across six dimensions, and usually says no. The two ideas I put through it came back 43 and 28.
It ships with its own validation harness: a dependency free suite that recomputes every score from its inputs, enforces an output contract per file, and fails the run when the arithmetic does not reconcile. The harness principle applied to a prompt system, which otherwise degrades quietly and never tells you.
npx find-me-saas init
If you have an agent that works in a demo and breaks in production, that is the conversation I enjoy most.