Guides · Series · 7 parts
Agentic AI for the Data Center Boom
In physical infrastructure, an agent’s autonomy is capped by reversibility and jurisdiction — not by how good the model is.
The authority gradient
Software agents fail cheaply. Re-run the job, roll back the commit, apologise for the bad summary. Data center agents do not get that. A cooling setpoint, a workload migration, a grid interaction, a buildout commitment — these touch physical systems where the undo button is expensive, slow, or absent.
That changes what determines an agent’s permitted autonomy. It is not model capability. It is how reversible the action is, who owns the reversal, and whether anyone outside the system is legally entitled to say no. Those constraints do not improve when the model does, which is why capability-led arguments for autonomy consistently overshoot in this domain.
The series builds one framework — the operational authority gradient — and then tests it against six real use cases, each of which caps autonomy for a different reason.
Why "self-healing" is the wrong goal
The uncomfortable finding running through the series is that the value is in diagnosis, not action. The MTTR you actually save is saved in understanding what went wrong, not in letting the agent act unsupervised — but the marketing sells the action, so that is what gets bought and built.
Each part isolates a different ceiling. Cooling is the most mature use case and the most misread. Workload placement acts only because the job can be moved, and whether it can be moved is someone else’s problem until suddenly it is the scheduler’s. Grid power is the outlier that caps autonomy on jurisdiction rather than physics, and an agent cannot measure a law. Buildout is where authority stops composing: two agents each acting inside their limits can commit a decision neither was allowed to make.
The final part is the one nobody has built. Six articles conclude that the boundary must be enforced outside the agent — so the last one asks what that outside is actually made of.
Which part solves my problem
The series reads in order, but if something is already broken, start here.
- “I need a defensible way to decide how much authority an agent gets.”Part 1 →
- “Everyone quotes the DeepMind cooling number at me and I need the real one.”Part 2 →
- “Leadership wants self-healing infrastructure and I am not sure that is the right ask.”Part 3 →
- “My scheduler moves workloads and I have not asked who owns the rollback.”Part 4 →
- “The agent is technically correct about the grid and still not allowed to act.”Part 5 →
- “Two agents each stayed within their limits and together made a decision neither could.”Part 6 →
- “We agreed on the boundaries and nothing actually enforces them.”Part 7 →
The series
7 of 7 published
Agentic AI in the Data Center: Operational Authority Gradient
Six use cases, one framework, and the uncomfortable reason 'self-healing' is the wrong goal.
Cooling Is Agentic AI's Most Mature Use Case - and Most Misread
Everyone quotes DeepMind's 40 percent. The number that actually matters is 30.
Self-Healing Infrastructure: Diagnosis, Not Action
The MTTR you save is in understanding, not action. The marketing sells you the action.
AI Workload Placement: Borrowed Reversibility, Not Owned
The scheduler gets to act because the job can be moved. Whether it can be moved is someone else's problem - until it's the scheduler's.
Data Center Power: Being Right Doesn't Grant Grid Authority
Every other use case caps autonomy on physics. This one caps it on jurisdiction - and the agent can't measure a law.
Authority Doesn't Compose: Why Coordinated Agents Need a Coordinator With Its Own Band
Two agents each acting inside their limits can commit a decision neither was allowed to make. The buildout is where that bites hardest.
AgentOps: A Band Is a Comment Until Something Enforces It
Six articles said enforce the boundary outside the agent. This is what the outside is made of - and almost nobody has built it.
Other guide series
Agentic AI with LangGraph
An eight-part guide to the part of agent engineering nobody demos: making the thing survive a restart, a concurrent write, a bad tool call, and a Tuesday deploy.
Harness Engineering
AI systems do not fail at the model layer. They fail in the code around it — and that code is a designable layer with its own architecture.
RAG Engineering
Most RAG failures happen upstream of the LLM, accumulate invisibly, and only surface as an incident. This series names each one.
Claude Code Playbook
Claude Code has four layers — context, skills, subagents, hooks — and the variable that decides which one you need is enforceability, not capability.
AI Control Plane
A fleet of agents needs a layer that is not itself an agent — observability, policy, orchestration, versioning, cost, and audit, built as infrastructure.
Codebase to Architecture Doc
A three-pass methodology for compressing any codebase into validated diagrams and decisions — and then a LangGraph pipeline that runs it for you.
