Agentic Engineering for Enterprise

Outline

1. What I've seen
2. Why the gains don't show
3. System 1: Knowledge corpus
4. System 2: Agentic SDLC
5. System 3: Validation harness
6. Putting it together

My experience

• Engineers moving slower than they could
• Confidence, context, validation: whichever is missing, they slow down
• Licences bought, usage invisible
• "Handed out the licences and shut our mouths"
• Review and QA become the bottleneck

Moving Too Fast

The AI Development Spectrum

Enterprise lives at the right end - or should

What Agentic Engineering Unlocks

• Pixel-perfect UI in seconds
• Productive in unfamiliar codebases day one
• Quality, security, architecture improvements
• Compressed timelines: 1mo -> few days

The New Engineering

Where are you?

Most rooms are at 2 or 3. This talk is how you get to 4.

Three Gaps

Three Systems

Each must be engineered

"We want the output to feel native to our ecosystem, not something that requires heavy rework to align with our platform conventions."

- software leader, large vehicle manufacturer
~200 engineers, 30 teams, 3 countries

A new senior hire

Same reason: doesn't know your platform

The Problem

• The Problem: AI doesn't know your platform
  • Standards, architecture, patterns - not followed
  • Templates, packages, services - not reused
• The Solution: a knowledge corpus as a build artifact

Why the obvious fixes fail

• Hand-written docs rot silently
• Search returns deprecated next to current
• A wiki has no contract
• Unreviewed AI docs make agents worse

Context engineering, enterprise edition

Three kinds of knowledge

The first two belong to the corpus. The third belongs to the SDLC.

Knowledge you want: curated

• Owned by a person - the enterprise architect
• Any vehicle: Confluence, repo, wiki
• Corpus ingests continuously, keeps it canonical
• Example: one error-handling practice, enterprise-wide

Knowledge of what is: generated

• Docs OF the packages, not API docs
• Validated before merge, stamped with the commit

Keeping it current

• Generated: mechanical staleness, "these 2 source files changed"
• Not calendar staleness, "this doc is 30 days old"
• Curated: an owner and a review cadence, wherever it lives

What else goes in

• Negative knowledge: when NOT to use it
• Honest gaps: every miss is logged, becomes backlog

How agents reach it

Agents ask for the doc they need. They don't load the platform.

Case Study: a corpus for a product platform

Every agent now knows how to use all of it

Every doc stamped to its commit. 0 failed checks out of 3,914.

The Problem

• The Problem: everyone starts from scratch
  • Best practices under-adopted, nothing shared, no oversight
  • Nobody can pick up anyone else's work
• The Solution: agree the rails, not the route
  • Where artifacts live, how context is shared, what gets checked
  • Don't invent them: start from Anthropic's AI-native SDLC

Where one engineer gets to

• Good. Not all the way
• Every engineer stops somewhere different
• Nothing persists, nothing is checked, nothing is shared

Anthropic's AI-native SDLC

• Every stage commits an artifact; the commit triggers the next stage
• A named human approves each gate
• Agents act up to the production gate, never through it

Intent, spec, plan

Each is a committed file. Nobody approves their own.

What you build around it

Four layers. The next slides take each in turn.

Plans cite the corpus

• Every platform decision cites a corpus doc
• No doc? The agent stops and says so
• A review layer audits the citations

Determinism

• Skills advise, hooks enforce
• The stage order is a state machine, not a convention
• Verification always runs, nobody has to remember

Parallel without collisions

• Worktrees - one per agent
• Skills that spin up, merge back, fix conflicts

Oversight and improvement engine

• The chain of commits is the audit trail
• Benchmark before you change a workflow
• Every incident becomes a regression eval

The Problem

• The Problem: AI is a fickle beast
  • Can't be trusted to check its own work
  • Can't be trusted not to break other things
  • Mocks its own code to force a pass
• The Solution: deterministic validation

Your harness sets your pace

The goal: it comes back right the first time, with evidence you can trust.

The layers

• Static: types, lint, format. Seconds. How conventions get enforced
• Unit: logic in-process, nothing external. Seconds
• Integration: real DB, real auth, ephemeral environment, 100% coverage. Minutes
• E2E: a real browser or simulator, clicking the real app
• Evals: only if you ship AI features

Best practices per layer

Integration: it runs the thing

Looks right is not works. 500s are found here, not by QA.

Visual verification

Advisory, never a gate

Case Study: a harness you can trust

Consumer app, iOS + Android, with AI features

How they connect

What it feels like

Less human time per build.
It should almost never come back busted.

The goal: work one layer up

• Whole epics go through the system
• You still drive; less intervention per feature
• Scale up as the harness holds

Baseline first

"We have 300 licenses out on GitHub... I heard 1,100 or something the other day. I didn't even know we had that many developers."

The head of AI enablement at a large US health insurer

• Measure DORA before you build anything
• Governance is what makes handing out licences safe

Key takeaways

It isn't the model

Sources

• The AI-native SDLC playbook, Anthropic. claude.com/resources/articles/the-ai-native-sdlc-playbook
• DORA: the four key metrics and the annual State of DevOps reports. dora.dev
• Model Context Protocol, the open standard agents use to reach tools and docs. modelcontextprotocol.io
• Claude Code hooks, how deterministic checks are enforced around an agent. docs.claude.com
• Case studies and quotes: my own client and product work, anonymised