Agentic DevOps
Put AI agents to work on a legacy estate - assess, map dependencies, containerize, and open governed pull requests - with the authorization model that lets you actually merge them.
4 weeks (2 sprints)
Who this is for
Companies sitting on a load-bearing 10-15+ year codebase that is the main brake on shipping. Teams that want AI leverage on the engineering process itself, not just AI features in the product. Agencies asked to modernize a legacy estate who need a defensible, governed method rather than a heroic rewrite.
When you need it
The biggest brake on AI adoption isn't model access - it's the 10-to-15-year-old codebase nobody fully understands anymore. Agents can now read that estate, map dependencies, containerize it, and open pull requests against it. The question stopped being 'can agents do this' and became 'who do you trust to point them at your production code, and how do you govern what they merge.'
What's delivered
- Estate and dependency map - in-scope legacy services inventoried, dependencies mapped (build-time and runtime), each ranked for modernization-readiness
- The governed agent pipeline - agents read the code, propose containerization and refactors, and open PRs, wired with scoped permissions, branch protection, human-in-the-loop merge gates, and a structured audit trail (WARRANT-aligned)
- Containerization pilot on 1-2 bounded services - dependencies resolved, Docker-containerized, agent-opened PRs reviewed, tested, and merged under the gate. Real merged output, not a deck
- Test and eval gates - characterization tests around legacy behaviour before it's touched, plus CI gates (build, test, security scan, behavioural diff) each agent PR must pass before a human reviews
- Authorization model for agent code changes - what agents may touch, under whose identity, with what blast radius, how it's logged. WARRANT-aligned vocabulary (read-only / side-effecting / identity-required / scoped)
- Sequenced modernization roadmap - the rest of the estate ordered by value, risk, and dependency, with the per-wave plan your team runs against the pipeline
- Board-ready summary - what was modernized, what the pilot proved, and the honest risk surface of pointing agents at production code
Why this consultant
Co-author of the WARRANT Standard for autonomous agent authorisation - an agent opening a PR against production is a side-effecting agent action, and authorising it properly (scoped identity, bounded blast radius, structured audit) is exactly what WARRANT specifies. Lived pattern, not theory: prismalOS runs adversarial eval swarms as CI/CD release gates - the same governance shape this sprint installs around agent-authored code. Plus twenty years operating legacy estates through scale, regulation, and three M&A processes including Bitstamp → Robinhood.
What this sprint is not
- A full migration delivery (1-2 pilot services end-to-end + the governed pipeline + the roadmap; the rest is a staffed programme, not a four-week sprint)
- Autonomous agents merging to main unsupervised (every agent PR passes automated gates and a human merge gate)
- 'AI will rewrite your app' (agents accelerate the mechanical work; sequencing and boundaries are the senior work)
- A product-AI engagement (that's the AI Integration Strategy Sprint)
Fixed price, fixed scope. Every engagement carries the four guarantees: Sprint 1 Escape Hatch, Board-Ready Quality, Three-Client Cap, Skin-in-the-Game Pledge.