Written, not implied
A spec an agent can run is also a spec a regulator can read. The two requirements turn out to be the same requirement.
How we deliver · Lifecycle · Methodology · Engagement
SECTION 04 / DELIVERY
Agents accelerate.
People approve.
AI agents do a great deal of the building now. That changes how fast the work moves — it does not change who is accountable for what ships.
The agentic delivery lifecycle
01
Six phases, six gates
Every phase is accelerated by agents and closed by a named human. Nothing advances because it looks finished; it advances because someone signed for it.
| Phase | What happens | Human gate |
|---|---|---|
| 01 · Discover | Use-case triage against value and feasibility; data audit; agents sweep your documents and systems to map what already exists. | BUSINESS CASE SIGNED |
| 02 · Design | Solution architecture, evaluation design, and guardrail specification — written so both agents and auditors can execute it. | ARCHITECTURE APPROVED |
| 03 · Build | Agent teams implement, test, and document in parallel; engineers direct the work and review every change. | EVAL SUITE PASSING |
| 04 · Evaluate | Offline evals, red-teaming, and user acceptance with your subject-matter experts on real data. | QUALITY BAR MET |
| 05 · Deploy | Staged rollout with human-in-the-loop first; integration hardening; operations training. | READINESS REVIEW |
| 06 · Operate | Monitoring, drift detection, retraining, and cost management, run as a service. | MONTHLY VALUE REVIEW |
Agentic methodology
02
How agentic development
actually works
Five rules that make agent-accelerated delivery safe enough to put in front of an auditor — and fast enough to be worth doing.
The spec is the contract. If a human cannot audit it and an agent
cannot execute it, it is not a specification — it is a wish.
- Spec-first. Every feature begins as a written specification an agent can execute and a human can audit.
- Agent teams. Coding agents implement, test, and document in parallel. Engineers set direction, review output, and own the result.
- Evals as contract. System behaviour is defined by evaluation suites that run on every change. Quality is measured, not asserted.
- Guardrails built in. Grounding, permissions, and output checks are part of the architecture from sheet one, not a patch before launch.
- Human gates. People approve what ships. Agents accelerate the work; they never deploy themselves.
Rule 01
Rule 02
Measured, not asserted
“It seems to work” is not a status. The eval suite is the status, and it runs on every change.
Rule 03
Grounded, not fluent
Retrieval and citation are structural. A confident answer with no source is a defect, not a feature.
Rule 04
Gated, not autonomous
Speed comes from the agents. Accountability comes from a named person at every phase boundary.
The loop
01Spec
02Agents build
03Evals run
04Human review
05Ship
06Observe & improve
Engagement model
03
Start small. Measure honestly.
Scale what works.
Every phase ends with something running and a number attached to it. We transfer knowledge as we go: your team owns the outcome, not a dependency on us.
| Phase | What we do together | You get | Typical timeline |
|---|---|---|---|
| Sprint zero | Use-case prioritisation workshop, data readiness assessment, and a value model for each candidate. | A ranked portfolio and a pilot plan | 2 WEEKS |
| Pilot | One use case built on your data against agreed KPIs — agents accelerate, your team stays in the loop throughout. | A working system and a measured result | 4–8 WEEKS |
| Production | Hardening, integration into your systems and workflows, security review, and operations training. | A deployed system your team runs | 8–12 WEEKS |
| Operate & scale | MLOps as a service — monitoring, retraining, cost control — while the next use cases enter the pipeline. | A compounding AI portfolio | ONGOING |
Watch The pilot’s KPI is agreed before a line of code is written. If we can’t name the number, we haven’t found the use case yet.
Working with us
04
Partner, not vendor
01
One team, end to end
The engineers who design the model also wire the devices, build the pipeline, and run it in production. No handoff, no seam for the work to fall through.
02
Knowledge transferred
We build internal capability as we go. The measure of a good engagement is that your team can run and extend what we built without calling us.
03
Numbers, not narratives
A baseline before, a measurement after, and a monthly value review once it is running. Everything else is decoration.
Next step
05
Two weeks to a
ranked portfolio
Sprint zero is small, fixed, and ends with a plan you could hand to another firm if you wanted to. That is rather the point.