Most firms added AI to an unchanged delivery model. We changed the model, on our own projects, for a year, before we sold a minute of it to anyone. Here is exactly how a pod is assembled, what the agents actually do, and the four things they are never allowed to touch.
Every pod is assembled after we know what you're building. Five senior people are drawn from a bench of 180 and matched to the work. The same four agent roles are on every pod; the specialists change with the work. When it's done, the pod comes apart.
Read it left to right.Read it top to bottom. Five people and a standing agent fleet are matched to the engagement by what the work actually needs. The seam between the two halves is where a human is pulled in. See §04.
They oversee and validate. They don't execute, and that's the whole economic argument. How much of each role you get is set by what you're building and how many pods it takes. You are buying senior judgment by the hour it's actually needed.
Owns the shape of the thing. Writes the definition of done that every agent iterates against, and is the only person who can change it.
Owns sequence, scope and your calendar. The person you call. Project management is the one thing we tried to automate and deliberately walked back.
Owns the standard the code is held to. Reviews what the agents produce, writes the parts that need a person, and does the final compilation.
Owns acceptance criteria and test strategy. Sets the bar the QC agent tests against, and signs off that done actually means done.
Owns risk. Reviews at design time, before a line is written, and again before release. On the pod from day one.
Five roles on every pod · allocation set per engagement · 180 engineers behind them
Four agent roles are on every pod. Specialists are added when the engagement needs them, and each specialist agent is supervised by someone on the bench who has done that job for real. An unsupervised specialist agent is just a confident stranger.
Flows, states, and the screens that carry them. Design decisions stay with the architect.
The build. Its brief is the definition of done, which it can check its own work against.
Environments, pipelines and configuration, everywhere below production.
Tests, regression, security scanning and the acceptance criteria the QC lead wrote.
Added for the engagement and retired with it.
"Done" is written at the start as a checkable definition: acceptance criteria, tests, security constraints. An agent keeps going until it passes or until it has to ask.
Steps 02 and 03 repeat without us. That loop is where the velocity comes from. Steps 01, 04 and 05 are people, every time.
The boundary is enforced by the environments themselves. Agents have real autonomy below the line, they can break things and sometimes do, and no credential that reaches production is ever in an agent's hands.
Agents create, destroy and rebuild freely. Nothing here is precious.
The QC agent runs the suite the QC lead defined. Failures loop back automatically.
Production-shaped. Where the human checkpoint happens and infosec does the pre-release review.
A person runs the release. Agents hold no production credentials and have no path to one.
Same rule whether it's our infrastructure or yours
These four never move, on any engagement, for any client. Where the agent's domain stops is not something we work out per project; it was settled early, after a year of testing where the technology fails, and it holds.
The build that becomes a release is produced by the senior developer.
Deployment to production is a human action with a human name attached.
Who can reach what, and under which rules, is decided and changed by people.
The infosec lead signs off before release. An agent can prepare the evidence; a person signs it.
A design-time review sets the constraints the definition of done has to carry: data flows, retention, what may touch a model. A pre-release review checks what was actually built against them. Both are agent-assisted and human-signed.
Agents run under a standing policy set that is the same on every engagement: no secrets in context, no network egress outside an allowlist, no writes outside the working repository, and every action logged and attributable to the session that took it. That is the floor; the full control set is part of the security review, and it doesn't get relaxed per client.
Certified in 2026, and audited by someone outside the company. Six of our clients are banks, insurers or public financial institutions, which is why the infosec lead joins the pod on day one and stays on it.
The pod is small on purpose. The company behind it runs to 180. When an engagement turns out to need a payments specialist or someone who has run a core banking migration, they already work here. Nearshore, same time zone, some of them past twenty years at Takeoff.
Uruguay · Argentina · Colombia · Ecuador. Your morning is our morning.