Agents don't read your dashboards.
Humans write specs. Agents do everything else. The guardrails we have were built for the old shape of the SDLC.
The SDLC just changed
Humans write specs. Agents do everything else. The guardrails we have were built for the old shape.
A couple of years ago the loop was human-written and human-reviewed: spec, write, review, CI, ship, with people at every step. Today the common shape is human-spec'd and agent-built: a human writes the spec, an agent writes the code, an agent installs the dependencies, an agent checks its own work, and a human ships it.
Security tooling hasn't moved with it. Dashboards, review queues, scanner reports: all of it assumes a person is looking. Agents don't read your dashboards. They need guardrails they can call inside their own loop.
Two incidents
axios@1.14.1: a patch release ships with a malicious transitive dependency. Any agent doing npm install axios auto-resolves it. There were stages where it could have been caught. None were watching.
The Ashley incident: an autonomous agent hit a 503, improvised, and posted internal system data to every organizer worldwide. The founder caught it manually.
Neither is a story about a bad model. The agent didn't have a bug; it had too much autonomy and no rail.
Guardrails for agents. Dashboards for humans.
Agents get guardrails: hooks, skills, and MCP tools the agent calls inside its own loop. No retraining, no policy file.
Agents verify their own work: a threat enumerated at design time becomes a check the same agent runs at code-gen time. One id, the design_review_id, threads them together.
Humans get a dashboard: auditors review reasoning, not a thousand PRs. The chain (spec, threat, guardrail, event, decision) is the artifact.
What ships today
Code-gen time, live now: a design review opened from your spec, a pre-write verifier running nine guardrails on every write and edit, and an install gate checking npm and pip packages against live registry signals and known CVEs. Every decision lands in an auditable lineage.
The layer above scanners and CI (orchestration, triage, the full threat graph) is the v1 roadmap, anchored to the same review id.
FlowRail is live as a POC and onboarding design partners. One command: npx @flowrail/init.