Same four phases every time. The only thing that changes is which workflow survives week one.
I sit with the people doing the work and write down every hand-off. What gets retyped, what gets chased, what gets forgotten. You get a scored list of candidates, and the ones I’d leave alone.
One workflow. Into your stack, not a sandbox. Evals before features: I need to prove it behaves before anyone trusts it with a customer.
A runbook a non-engineer can follow, a walkthrough with whoever owns it, and the keys. No dependency on me, by design.
One hour, thirty days later. What drifted, what the team worked around, what earned the next build.
If I can’t measure it behaving, it doesn’t go near a customer.
A portfolio of half-built agents is worse than one that runs.
The build isn’t done until someone on your team can run it without me.
Most of the value is in plumbing nobody wants to demo.
Forty-five minutes, no deck. If there’s nothing worth automating, I’ll say so.
