Insights
Field notes from inside the boundary
Agents, identity, and shipping under federal accreditation. Commentary from 20 years of mission engineering across DoD, DHA, SSA, and CMS.
Latest
RSS feed →-
Dayel Ostraco
Anatomy of a Governed AI Factory
The Colophon case study made the claim and the routing essay made the model calls; this is the architecture that makes a governed AI factory real and auditable, from the coordination plane to the eval loop.
- ai
- agents
- architecture
- governance
- evals
- federal
-
Dayel Ostraco
We Don't Have a Favorite Model. We Have a Routing Table.
The question I get is which model is best. It is the wrong question. In our governed federal factory the first filter is the boundary, the second is failure mode, and the eval that matters is disagreement. The model is a commodity. The routing is the craft.
- ai
- agents
- llm
- multi-model
- evals
- federal
-
Dayel Ostraco
Compliance Is a Byproduct, Not a Phase.
Federal software treats compliance as a phase you survive at the end. I built Colophon to invert that: a governed AI agent bench that emits the RMF artifact set as a byproduct of every shipped increment. Humans decide, the bench ships.
- ai
- agents
- ato
- federal
- compliance
- rmf
- colophon
-
Dayel Ostraco
Swap the Model. Keep the ATO.
Frontier models version monthly; accreditation does not. Fuse the agent's identity to a model name and every release forces a significance question you cannot answer cheaply. Abstract the model behind a stable identity and a swap becomes a change request with a regression gate.
- ai
- agents
- ato
- rmf
- federal
- procurement
-
Dayel Ostraco
Least Privilege Is a List, Not a Prompt.
A prompt that says do not drop the database is a request. A deny-by-default command allow-list, enforced beneath the agent, is a control. What the list looks like in practice, where it must be enforced, and the escalation paths it forecloses.
- ai
- agents
- cyber
- least-privilege
- nist
- federal
-
Dayel Ostraco
The Command Line Is an API Now. Govern It Accordingly.
The command line was only ever safe because a human was holding it. With an agent at the keyboard, the CLI needs everything an API already has: a stable contract, structured output, per-call authorization, and audit. Here is the gap, and how I close it.
- ai
- agents
- cli
- api
- governance
- federal
-
Dayel Ostraco
Your Agents Are Accounts. The Boundary Hasn't Moved.
Every agent-platform pitch claims your IAM cannot hold agents. I keep asking which NIST 800-53 control stops working when the account belongs to an agent, and nobody answers. Agents are accounts. The real work is governing them at machine speed.
- ai
- agents
- ato
- federal
- identity
- iam