No accountability model
Put named agents under explicit human input and approval contracts.
Inspect the HITL mechanics01 / NATIVE AGENT RUNTIME
AGENT AUTOMATIONUNDER HUMAN COMMAND.
Built for platform and AI engineering teams that answer to security reviews, auditors, and their own infrastructure.
- SINGLE BINARY
- MIT
- POSTGRES HITL
- SBOM + SIGNED
02 / WHERE PILOTS STALL
THE HARD PART STARTS AFTER THE DEMO.
Agent pilots do not fail because the demo was too small. They fail when control, deployment, and evidence arrive too late.
Dependency swamp
Ship one Rust binary. No Python or Node runtime is required in production.
Inspect the releaseProcurement wall
Bring checksums, an SBOM, signatures, source, and an on-prem deployment path.
Review the posture03 / THE OPERATING MODEL
CONTROL IS PART OF THE RUNTIME.
Every operating claim points to source, documentation, or a release artifact.
Human chain of command
Agents ask for missing context. Consequential tools wait for a separate, fail-closed approval.
HITL contractDeploys where you are allowed
A native binary runs with your provider, your network controls, and PostgreSQL when scale needs it.
Deployment guideSurvives procurement
MIT source, CycloneDX SBOM, keyless signatures, checksums, and a signed OCI path.
Release evidenceOperates like infrastructure
Replicas coordinate durable ownership, replay, encrypted human answers, and graceful drain.
Multi-replica contractNo platform bet
Switch between native and OpenAI-compatible providers through configuration, including local Ollama.
Provider matrix04 / WHAT YOUR TEAM SHIPS
A CREW YOU CAN READ. A RUNTIME YOU CAN OPERATE.
Define the team in readable Lua. The native runtime owns scheduling, policy, providers, and durable execution.
local crew = Crew.new({
goal = "Assess release risk",
provider = "openai",
model = "gpt-5.6-luna",
require_approval = { "file_write" },
})
crew:add_agent(Agent.new({
name = "researcher",
goal = "Collect release evidence",
tools = { "ask_human" },
}))
crew:add_agent(Agent.new({
name = "reviewer",
goal = "Challenge evidence and decide",
tools = { "file_write" },
}))
crew:add_task({ name = "collect", description = "List the riskiest changes in this release." })
crew:add_task({
name = "decide",
agent = "reviewer",
depends_on = { "collect" },
description = "Write a go/no-go plan for: ${results.collect.output}",
})
crew:run()COORDINATOR
Let one agent delegate bounded work to specialists — callable as agent__name tools or whole sub-flows.
Run agent-as-toolROUNDTABLE
Adversaries debate under forced falsification — every turn ends with what would prove it wrong — then a moderator synthesizes to a schema.
Run the stock debateV3.0.0 / RELEASE EVIDENCE
- CYCLONEDX SBOM
- KEYLESS SIGNATURES
- SIGNED OCI
- MIT
MAKE THE OPERATING MODEL EXPLICIT