Cost Guard
Real-time solvency gating. Halts execution when projected spend exceeds budget. Per-tier rate cards with configurable thresholds.
ag-os check 1.0 --tier gpu_a100
The governance kernel for AI agents.
Cost limits, policy enforcement, and audit trails. Built into your agent runtime — not a SaaS, not a sidecar, no lock-in.
9
Constitution rules
6
Provider surfaces
0cloud
Lock-in dependencies
The problem
Autonomous agents are powerful. Ungoverned agents are dangerous.
Agentic pipelines introduce unbounded state transitions into production. The operational failure modes group into three observables.
-
01
Unbounded cost
Execution loops retry without solvency checks. A single agent can exhaust thousands in GPU compute before any human notices.
Observed: $4,200 burned in a 12 h window -
02
State opacity
Transitions occur in transient memory contexts. Post-incident forensics lack a deterministic replay mechanism.
Audit coverage: 0% in baseline runs -
03
Static competence
Agents discard execution context between sessions. Identical failure patterns repeat across independent runs.
Retry storms: 12 cycles avg before escalation
The architecture
Six modules. One install. Complete governance.
Independent abstractions intercept agent state transitions without modifying the underlying logic. Each module is swappable; every interface is documented.
The Constitution
Nine governance rules enforced at every state transition. Policy-as-code defined in Markdown. Extensible by design.
.agent/rules/08-economic-safety.md
Flight Recorder
Append-only state machine with deterministic replay. SQLite WAL-backed. Every operation traced end-to-end.
ag-os status --tail 100
DreamEngine
Self-improvement loop. Scans friction logs, identifies failure patterns, synthesizes governance patches automatically.
ag-os dream --apply
Provider Registry
Six swappable surfaces — secrets, issues, cost, state, telemetry, policy. Works offline. Scales to cloud.
providers: { state: redis }
MCP Server
Exposes governance as Model Context Protocol tools. Any MCP-compatible agent gets governance built-in.
ag-os serve --transport stdio
The evidence
Measured impact on a production pipeline.
A 7-agent content pipeline operating across 5 regional tenants. Every metric reflects the kernel's enforcement, not advertising.
| Metric | Baseline | Governed |
|---|---|---|
| Budget overrun | Uncapped | $0.00 |
| Avg retry cycles | 12.0 | 2.1 |
| Audit coverage | 0% | 100% |
| Self-healed patches | — | 6 generated |
| Policy violations | untracked | 14 caught |
$pip install ag-os
✓ ag-os 1.4.0 installed
$ag-os init --governance
✓ 9 Constitution rules active
$ag-os run pipeline.yaml
▸ executing 7-agent pipeline…
step 1 · DataFetcher · ok · $0.02
step 2 · TextAnalyzer · ok · $0.15
step 3 · Synthesizer · ok · $0.33
✓ pipeline complete · total $0.50
$ag-os dream
✓ 2 governance patches synthesized
$
Reference implementation
Governing a multi-agent content pipeline.
A real-world deployment running under Antigravity OS — what's being governed, how the kernel applies, and what was measured.
The system
What's being governed
Multi-tenant content pipeline with seven specialized agents, each with sub-agent decomposition.
- 7 agents, 15+ sub-agents
- 5 regional tenants (IL, US, MX, ES, PT)
- GPU-accelerated inference endpoints
Modules active
Governance applied
Four kernel modules govern this pipeline end-to-end — cost, policy, trace, and self-improvement.
- Cost Guard · per-operation solvency
- Constitution · 9 rules enforced
- Flight Recorder · full trace
- DreamEngine · patch synthesis
Observed delta
Measured impact
Operational metrics after integrating the kernel into the pipeline's execution loop.
- Budget overrun: $0
- Retry storms: 12 → 2.1
- Audit coverage: 0% → 100%
- Self-healed: 6 patches
Get started
Install the kernel in seconds.
One command. Six modules. Zero cloud dependencies. Read the source, fork the repo, ship governed agents today.