Training · ComposerMoveemergingpartial

Center-of-Excellence AI-Native Engineering

also known as CoE agent training, AI CoE builder program, AI-native engineering team

Craft Path: ComposerOrchestrator

An enterprise Center of Excellence dedicated to training internal builders to design, build, and run agents on the company's own stack. Vendor technical-enablement partners support the CoE. The goal is a standing agent-building capacity across the organisation, not one-off experiments.

How the learner advances

flowchart TD A[Executive mandate + vendor enablement partner] --> B[Stand up CoE team 5-20 engineers with charter] B --> C[Rolling cohort embeds 4-12 weeks per cohort] C --> D[Build real agent against home-team problem] D --> E[CoE + vendor architecture review + failure-mode audit] E --> F{Agent shipped before cohort exit?} F -- No --> D F -- Yes --> G[Graduate returns as home-team expert] G --> H[CoE publishes reference architecture + prompt library] H --> I[Measure business outcomes report to executive sponsor] I --> C

Intent. Build a standing internal capacity to design, deploy, and operate agents at enterprise scale by training cohorts of engineers inside a dedicated CoE with vendor or specialist technical-enablement support.

When to apply. Use this move when an organisation has passed the proof-of-concept stage and needs repeatable agent-building capability across multiple teams — not when the first prototype is still being built. It requires an executive mandate, a dedicated team of 5–20 engineers, and at least one vendor or specialist partner willing to provide deep technical enablement, not just account management.

Threshold — earns the next step. CoE graduates have each shipped a monitored production agent with a documented business outcome, and the internal reference architecture has been adopted by at least one team outside the CoE.

Masterpiece — the artifact that proves it. A portfolio of production agents built by CoE cohort graduates — each with a documented business outcome and monitored in production — plus an internal reference architecture standard adopted by teams outside the CoE, demonstrating that the organisation now has standing agent-building capacity rather than isolated expertise.

Example scenario

A Japanese technology integrator with 40,000 engineers is running scattered AI experiments across 30 business units but cannot get any agent past the proof-of-concept stage. Each team that builds something hits the same architecture problems, solves them independently, and cannot share the solution because there is no common platform. The CTO secures a board mandate: stand up a CoE with a two-year runway and a goal of 50 production agents shipped by end of year one. The company engages Anthropic as the technical-enablement partner. Anthropic provides three things. First, architecture office hours with CoE engineers every two weeks. Second, a code review of every agent before it moves to staging. Third, a failure-mode review checklist tailored to the company's security and compliance constraints. The first 12-engineer CoE cohort runs for eight weeks. Each engineer brings a real problem from their home team. By week six, all 12 have agents in staging. By week eight, nine are in production — the other three are blocked by data-access approvals outside the CoE's control, not by engineering failure. The nine who ship document their architectures in a shared internal wiki. By month three, the wiki holds enough patterns for the CoE to publish a reference architecture. It covers stack choices: Claude API, LangGraph, internal auth, and Datadog for observability. The CoE also publishes a prompt template library for the most common internal tasks and an evaluation harness that runs 50 standard test cases against any new agent before deployment. Cohort two's ramp time is 40% shorter because they start from the reference architecture rather than blank files. By the end of year one, 47 agents are in production across 18 business units. The masterpiece is not any individual agent but the portfolio. It contains 47 agents, each with a documented outcome, owned by engineers in business units who built them without CoE dependency. The reference architecture is what made cohort two's ramp 40% shorter than cohort one's.

Facets

Container — embedded
Mode — hands-on-buildmentored-unblockbyo-problem
Reach — org
Persona — buildermanager-leader
Craft (AI Fluency) — delegationdescriptiondiscernmentdiligence

Inputs

Executive mandate and CoE team — Leadership commitment to fund a dedicated team of 5–20 engineers for a sustained period, with a mandate to build production agents — not to evaluate them.
Vendor or specialist technical-enablement partner — An AI provider (such as Anthropic, OpenAI, Microsoft, or Google) or specialist consultancy that provides architecture guidance, code review, and failure-mode reviews to the CoE team — not just sales support or generic documentation.
Real production problems from internal stakeholders — A backlog of real automation opportunities from business units across the organisation — the raw material that each CoE cohort builds against.
Internal platform: reference architectures, prompt libraries, evaluation harnesses — A growing internal knowledge base that the CoE builds and maintains — reusable components that successive cohorts extend rather than rebuild from scratch.

Outputs

A more capable learner — Each CoE cohort graduate who has built and shipped a production agent — and who returns to their home team as a local agent-building expert able to mentor colleagues without CoE support.
Masterpiece: a portfolio of production agents with measurable outcomes — A growing portfolio of agents built by CoE graduates, each with a documented business outcome — tasks automated, hours freed, cost reduced — that constitutes the organisation's evidence base for continued investment.
Internal reference architecture — A documented, tested, organisation-specific agent architecture standard — covering stack choices, security model, data access patterns, and observability — adopted by teams outside the CoE.
Distributed agent-building experts — CoE-trained engineers embedded in business units across the organisation who can scope, build, and operate agents without returning to the CoE for every new project.

Steps (4)

Stand up the CoE with a mandate and partner
Secure executive sponsorship with a clear charter: the CoE produces production agents, not evaluations or strategy documents. Engage the technical-enablement partner and agree on what support they will provide — architecture reviews, office hours, failure-mode audits. Staff the CoE with engineers who have or can rapidly acquire foundational agent-building skills.
Run rolling cohort embeds
New builders from across the organisation embed with the CoE for 4–12 weeks. Each cohort member builds a real agent against a problem from their home team under CoE mentorship. The CoE lead reviews architecture decisions; the technical-enablement partner reviews for security, scalability, and failure modes. Cohort members ship before they exit.
Build and publish internal reference assets
As each cohort runs, the CoE documents what it learns. This produces a reference architecture for the company's stack, a prompt library with versioned templates, and an evaluation harness that tests agents against a common quality standard. These assets accumulate over time and reduce the ramp-up cost for each successive cohort.
Measure and report business outcomes
Track the agents shipped by CoE graduates: what they automate, what they save, what quality improvements they produce. Report this to the executive sponsor. The CoE's continued mandate depends on this evidence; without it, the programme is indistinguishable from an expensive experiment.

Principles

The CoE builds, not advises — a CoE that only produces strategy documents and internal presentations is not a builder programme.
Cohort graduates return as experts — the exit condition is not a certification but the ability to mentor a colleague in the home team without CoE support.
The vendor partner enables, not replaces — technical enablement from a vendor supplements CoE capability; the CoE retains architectural ownership.
Outcomes trump activity — the programme's evidence is production agents with measured business impact, not hours of training delivered or courses completed.

Unlocks methodologies (4)

A learner who completes this pattern is equipped to execute these methodology families:

Agent Construction Multi-Agent Design Coordination Evaluation

Known uses (2)

Known failure modes (3)

[strategy-document-coe]
The anti-pattern of a CoE that produces guidelines, evaluation reports, and governance frameworks but does not ship production agents. An advisory CoE builds institutional authority but not institutional capability; the builder programme threshold is not met.
[vendor-dependency-lock-in]
The anti-pattern of a CoE whose entire architecture expertise resides in the vendor technical-enablement team rather than in the CoE engineers. When the vendor relationship changes, the CoE's capability evaporates. Technical enablement must transfer knowledge, not just solve problems.
[cohort-without-graduation]
The anti-pattern of cohort members who embed with the CoE but do not ship an agent before returning to their home team. Without a shipped agent as the exit condition, the embed produces exposure but not capability, and the distributed-expert goal is not reached.

Related trainings (3)

Sources (2)

Provenance

Ecosystem: in-house
Added to catalog: 2026-05-27
Last updated: 2026-05-27
Verification status: partial

How the learner advances

Steps (4)

Stand up the CoE with a mandate and partner

Run rolling cohort embeds

Build and publish internal reference assets

Measure and report business outcomes