Every capability in the governed runtime, and what we're improving. This log reflects the current public release — grouped by area, each entry tied to a capability that ships today, with no fabricated version history.
Grouped by area · Shipped vs. Improved · links to the platform
Author an agent and every run flows through the same fail-closed governance chain — identity, budget, guardrails, registry, control-tower, policy, oversight, audit. Deny is the default.
Author rules that evaluate in priority order with the most-restrictive effect winning (deny > require_approval > allow), preview a decision and its per-rule trace before shipping, and gate changes against a golden-test suite.
Every action runs through dry-run → propose → approve/deny → execute → compensate with an immutable invocation ledger and an evidence pack. High-risk actions land in a human approval inbox; declared rollback keys enable compensation.
Set per-agent autonomy from suggest_only through autonomous. Oversight can only tighten a decision, never weaken the action's own approval floor — and admin break-glass force-executes with a mandatory audited reason.
Agents carry their own identity, allowed tools, allowed models, allowed data, and deployment channel. The runtime refuses to execute missing, draft, unpublished, cross-tenant, expired, paused, or unauthorized agents.
MCP servers register as pending and cannot be invoked until approved; block is an instant kill-switch. Governed invokes run a DLP scan on input, enforce a per-server tool allow-list, and record every call with redacted I/O.
An aggregated per-tenant view of agents, runs, spend, tools, and active controls — with one-click pause and tool-disable that the runtime enforces (a paused agent is rejected 409 AGENT_PAUSED).
Every governed outcome is recorded so a third party can check it. The audit ledger is hash-chained, outcomes emit signed receipts, and a provenance graph stitches the chain from human to outcome.
Every audit event is hash-chained so any insertion, edit, or deletion is detectable, with a verify endpoint that proves integrity over a range and returns the broken sequence if the chain was altered.
Each governed outcome — an agent run, an action execution, an approval decision — emits a compact, detached-signature receipt a third party can validate offline, with the signing key operator-supplied and fail-closed in production.
A queryable lineage graph stitches the chain Human → Agent → Skill → Prompt → Policy → Model → Tool → Artifact → Outcome → Approval, and every surfaced fact links back to its source filing, page, and box.
Author governed agents in plain English or code, then move them through a real lifecycle — draft, test, review, publish — with a reliability bar that decides what reaches production.
Author agents with skills, knowledge, guardrails, and model policy, then walk the draft → test → review → approve → publish → deploy → pause → retire lifecycle — with prompt and instruction versioning and a built-in evaluation harness.
A 0–100 score blends success rate, eval pass rate, and incidents over a rolling window. The registry refuses to publish an agent that has enough signal and scores below the bar (409 RELIABILITY_TOO_LOW) — while new agents publish freely and a scoring outage fails open.
Model object types, relationships, and verified metrics with object-, property-, and action-level permissions — plus ontology versioning, promotion, rollback, and migration so the model evolves safely.
Orchestrate multi-step work and let events trigger governed actions — without ever stepping outside the gates.
Compose multi-step workflows and event-condition-action automations that propose governed actions, run on a schedule or a queue, and stream through the real-time hub — every step still passing the same gates.
You can't govern what you can't measure. Per-run quality scoring, eval gates, a simulation lab, and a full-trace debugger turn agent quality into a number you can watch and gate on.
Each governed run is scored for quality — faithfulness, citation coverage, safety — with the trace captured for replay; gold eval suites act as release gates that block a regression before launch. Live runs return quality { overall: 100 }.
The debugger reconstructs an entire run — prompt → context → model → tools → approval → output — each step stamped with cost, latency, and verdict, and every hop linking back to the trust ledger.
A/B prompts and models against real traffic and compare the results before promoting a change — so a candidate proves itself before it touches production runs.
Put a budget on the runtime and tie every dollar to the work it bought.
Set tenant- or agent-scoped budgets with an alert threshold and an optional hard cap. Spend is computed live over a rolling 30-day window, and a run is blocked once projected spend would exceed a hard cap — emitting BudgetExceeded events.
Govern which models an agent may use and route runs through a compute broker — so model choice is a policy decision, with usage, adoption, and ROI rolling up alongside cost.
Bundle a tenant's governed solution into a versioned, signed, portable artifact you can verify and re-deploy — across tenants and air-gapped environments.
A pack bundles ontology types, policies, actions, and agent metadata; its contents are canonicalized, hashed, and HMAC-signed. verifyPack checks both the content hash and the signature — any tamper breaks hashOk, any wrong secret breaks signatureOk.
Importing a pack shows a per-section diff — added, changed, unchanged — against your live definitions, then verifies before it applies and refuses an invalid pack (409). Transfer happens through the artifact, never a direct cross-tenant write.
Read how the governed runtime fits together, or see it run against your own work.