Field notes
Why the substrate is the bottleneck
The frontier of agent capability is no longer the model — it is the infrastructure the model has to drive. Here is how we think about the runtime layer that closes the gap.
Most agent failures people blame on the model are actually substrate failures. The model wanted to run a command, but the runtime was not built to be driven by software, so the action stalled in a human approval queue.
When we started NoHumanInfra we kept asking the same question: what would infrastructure look like if it were designed to be operated by software first, and by people second?
Three properties
The answer kept collapsing into three properties that legacy infrastructure refuses to give agents:
- Provisioning on demand — compute appears when policy allows it, with no ticket and no waiting on an operator.
- Verifiable actions — every deploy, tool call, and state mutation is signed and replayable, so a system can prove what happened.
- Safe degradation — when something goes wrong, the system falls back to a read-only route with deterministic recovery instead of paging a human.
Each of these is a substrate property. None of them come from a smarter model.
What changes
Once the substrate can be trusted to act, the agent loop gets dramatically shorter. A fleet that used to file a ticket, wait, get paged, and roll back can now provision, verify, and roll forward inside a single bounded execution — and hand a human a signed receipt at the end.
That is the bet we are making. We will be writing more of these notes as the runtime grows — on verification, on recovery, and on what it takes to run production workloads no human is staffed to watch.