How we build AI agents that get installed, used, and paid back. Eight phases. Two hard gates. Ninety-three steps. The operating discipline behind every Easton build.
~6 hours · blocks Build until client sign-off
The first phase has one job: produce a written PRD the client signs off before any code gets written. No PRD, no Build. It's the cheapest place to catch a misunderstanding, and the most expensive place to discover one is week three.
Easton interviews the SME who owns the workflow, maps it step by step, and collects five to ten golden examples (real input, real expected output) so QA later has a fixed bar to test against.
~4 hours · prep work before the build starts
Four hours go into the boring decisions that wreck projects when they're skipped. We pick specific tools inside the chosen category, draw the agent flow, and request every credential and API key formally so nothing dies in a Slack DM.
Token cost gets estimated against the golden set so the expected monthly run cost is grounded in maths, not vibes. Cost alerts get wired before a single line of agent code ships.
~20 hours · only starts once Phase 1 is signed off
The part most agencies sell as the whole project. At Easton it's roughly forty percent of the work, and it never starts without a signed PRD upstream.
Happy path first, tested against golden examples. Error handling, retries, and structured logs are first-class citizens, not afterthoughts. Hours get logged daily so a budget overrun shows up while there's still time to do something about it.
~8 hours · everything Easton can break before the client sees it
Before the client opens the agent, we break it. Network pulled. Bad input shoved in. Auth tokens expired. Rate limits pushed. A 100-run load test confirms cost per run lines up with the estimate.
Client UAT doesn't open until zero P0/P1 bugs are outstanding. Anything looser turns UAT into bug triage instead of the trust-building session it's supposed to be.
~6 hours · blocks Deploy until written sign-off
The agent stands up in staging. You get a live walkthrough, run your own tests against real data, and mark each success criterion pass or fail on the spot.
Production deploy doesn't happen without written sign-off. If sign-off is blocked, the blocker gets named in writing and the project sits in UAT until it's cleared. No silent slips into prod.
~4 hours · only starts once Phase 5 is signed off
Deploy is fast because the work behind it is done. Production goes live, failure alerts get wired to you and Easton, and a forced test failure confirms the alert actually fires before anyone relies on it.
A 10–15 minute Loom walkthrough plus a one-page runbook ship the same day. Optional 30-minute Q&A if the build warrants one. Nobody disappears.
2 weeks · bug-only window, included in the price
The two weeks after deploy are where most projects quietly go sideways elsewhere. We scope that window into every build. Logs reviewed daily for the first three days, weekly thereafter. Every alert triaged the same day.
Bug-fixes only. Feature requests get captured as new project candidates and quoted properly, so the live agent stays stable.
~3 hours · locks the project and feeds the next one
The final phase locks hours, profitability, and learnings. The internal retro names what was estimated vs actual, what broke, and what's worth lifting into the pattern library for the next build.
Skipping this is how an agency forgets what blew up and repeats it on the next project. Easton runs it on every build, every time.
Most of the eight phases are visual order. Two of them are non-negotiable. These gates protect every build from the common patterns that kill AI projects elsewhere.
A 90-minute conversation in week one is worth a 90-hour rewrite in week six. We won't move into Build until the PRD has named the goal, inputs, outputs, success criteria, and the out-of-scope list, and the client has signed off in writing.
An agent that demos well in staging isn't an installed system. We hold production deploy until the client has run the agent against real data, marked every success criterion pass or fail, and confirmed in writing they're ready to go live.
Book a 30-minute discovery call. You'll leave with a scoping note inside 48 hours that names the agent, the build hours, and the expected monthly running cost.