Agentic Delivery Architecture

The squad is getting smaller. The system around it cannot remain manual.

Traditional delivery divided work across a stack of specialists. Agile brought those specialists into a two-pizza team. Agentic delivery compresses the team again: a domain authority and a Forward Platform Engineer, supported by agents and a shared enterprise delivery system.

The challenge is no longer how to staff a squad. It is how to engineer the architecture, controls, support and organisational memory that let a two-person cell own an outcome safely.

Book a call See the operating model

The squad evolution

From a stack of specialists, to a persistent team, to a smaller team using agents, to a domain authority and a Forward Platform Engineer each orchestrating specialist agents. Five people are one point on an evolving spectrum, not the final canonical shape.

The squad evolution in four stages. 1, Traditional stack: work passes through specialist layers, business, architecture, engineering, test, release, operations, with many hand-offs and slow feedback. 2, Two-pizza squad: a persistent cross-functional team with broader ownership but still coordination-heavy. 3, Hybrid squad: a smaller human team using AI inside existing delivery. 4, One-to-one agentic cell: a domain authority and a Forward Platform Engineer amplified by agents, outcome-owned and evidence-led. What changes: team size shrinks, capability becomes reusable, controls move into the platform, and the workflow becomes the unit of transformation.

Traditional stack

Work passes through specialist layers. Many hand-offs, slow feedback.

Two-pizza squad

A persistent cross-functional team. Broader ownership, but the coordination burden remains.

Hybrid agentic squad

A smaller human team uses agents inside existing delivery processes.

Agentic cell

One domain authority and one Forward Platform Engineer, each orchestrating specialist agents, drawing on shared enterprise capabilities when the work requires them.

The 1:1 model will not suit every production service, but it becomes the default starting shape for bounded workflows and Domain Building Blocks. Extra people are pulled in because the work needs them, not because a seat must be filled.

The system around the 1:1 agentic cell

The squad can shrink only because the organisation around it becomes more engineered. The cell sits inside a paved road, wrapped by an accountability backbone. This is not AI-amplified delivery. It is a unit of accountable autonomy.

Three concentric layers around the one-to-one agentic cell. At the centre, a domain authority and a Forward Platform Engineer, with agents amplifying the work. Around them, the agentic paved road: data and context, reusable skills and Domain Building Blocks, cost and run-rate, security and policy, code and deployment, evals and testing, observability, and tools and connectors. Around that, the accountability backbone: agent contracts, named executive ownership, reference architecture, audit trail, mandates and permissions, TDRs, the trust ledger, and human override. Why this matters: safe autonomy, repeatable delivery, supportable production, and knowledge that stays with the client.

1. Agentic paved road

A pre-integrated route from idea to production: approved models, tools and connectors; reusable agent skills and Domain Building Blocks; isolated environments; automated testing, evaluation and policy checks; deployment, observability and cost metering; versioned prompts, models, data, code and agent configurations. Without it, every cell rebuilds a fragile local stack.

2. Trust and control plane

Controls travel with the work, not through meetings: agent identity and Know Your Agent; machine-readable Agent Contracts; mandates, permissions and operating boundaries; TDRs for material decisions; evidence, outcomes and model versions written to the trust ledger; revocation, rollback and kill switches.

3. Production support

A two-person cell cannot recreate a bank's operational machinery. It draws on shared service ownership and on-call; incident, problem and change pathways designed for agents; resilience and recovery testing; agent failure and degradation patterns; human takeover and exception handling; common operational telemetry.

4. Enterprise capability mesh

Architecture, data, cyber, risk, legal, compliance and procurement move from permanent seats and committee gates into reusable services: controls as code; architecture fitness functions; approved data products; reusable contractual and regulatory patterns; automated assurance; specialists on call for exceptions and novel decisions.

5. Workforce and learning pipeline

A senior 1:1 model must create the next generation, not consume expertise: rotating apprentices and emerging engineers; recorded decisions and reusable agent skills; structured shadowing of domain experts; progression from operator to orchestrator to accountable owner; deliberate transfer of supplier knowledge into the client.

The centre gets smaller, and more capable

It owns the paved road, the reference architecture and the trust ledger. It does not own the squad's backlog.

From discovery to scale

How an agentic cell becomes a production operating model. The engagement moves from watching the real work to a repeatable portfolio of small, accountable cells.

Six stages from discovery to scale, sitting beneath shared enterprise services (architecture, data, risk, security, procurement, HR, operations). 1, Shadow: sit with the people doing the work. 2, Map: understand the workflow, friction and decisions. 3, Build: create the first agentic workflow against real systems. 4, Validate: test with users, controls and edge cases. 5, Ship: deploy with evidence, support and rollback. 6, Scale: turn one working cell into a repeatable model. What must be solved: support, observability, funding, skills, governance and institutional memory. End state: small accountable cells, a reusable paved road, evidence over status, and scaled organisational learning.

Shadow, Map

Sit with the people doing the work. Understand the workflow, its friction and the decisions inside it.

Build, Validate

Create the first agentic workflow against real systems. Test it with users, controls and edge cases.

Ship, Scale

Deploy with evidence, support and rollback. Turn one working cell into a repeatable model. The wider path is the operating model.

The four questions this answers

Agentic Delivery Architecture, not merely agentic squads. The offer answers the questions an accountable executive actually has.

The smallest unit

What is the smallest accountable delivery unit for our organisation?

Embed or share

Which capabilities must live in the cell, and which should become shared platform services?

Operate at scale

How do we operate, support and assure hundreds of small cells without recreating central bureaucracy?

Knowledge and talent

How do we preserve institutional knowledge and grow talent as human teams become smaller?

Beyond the demonstration

Uber has publicly described the forward-deployed pair as a discovery mechanism: an AI-proficient engineer beside a domain expert, observing the real work, redesigning the workflow and shipping in weeks. That validates the shape. It describes the intervention, not yet the complete production operating model.

The unanswered question is what must become a shared organisational capability so that a two-person cell can safely own a production outcome. That is the question this page answers. Uber shows how to discover and automate a workflow. DTA describes how the resulting delivery model becomes governable, supportable and repeatable across a regulated enterprise.

Bring us a challenge worth solving.

Tell us the proposition, architecture decision, or transformation problem you are facing. We will tell you, honestly, whether and how we can help.

Discuss a challenge

Get the Observatory, our living models of banking, by email

Living models and Field Notes on the future of banking. Published when the model changes, not on a schedule.