Agent actions need a common language.
Why we're building an open protocol for governing autonomous actions, and why OSuite came second.
Research, protocol, product
Agent actions need a common language.
Ond is developing Ambit, an open protocol for describing, authorizing and proving what AI agents do across runtimes. OSuite is how we make it practical, and how we test it.
Backed by
Official partner
Official partner
The path
From a research question to an open protocol.
Each step is published, and each one builds on the last. The research defines the action; Ambit makes it shared; OSuite puts it to work.
-
01Principle
Final Authority
Final Authority in AI Governance
Authority over high-impact actions belongs to the organization that bears the consequences, not the model provider.
-
02Research
PCAA
Proof-Carrying Agent Actions
Govern the action, not the model: one certificate that carries its approval and proof across any runtime.
-
03Research
CAVA
Canonical Action Verification and Attestation
One action, one identity, whether it arrives as a shell command, a tool call or an API request.
-
04Research
ZeroGate
Trust-preserving fast paths
Approve ahead, check at the boundary in about 10 milliseconds.
-
05Research
Matoi
Selective revalidation of consequence evidence
Reuse evidence. Recheck authority.
-
06Protocol
Ambit
First layer: Action Effect Spec
An open protocol for agent actions. Its effects layer and test suite are public today.
-
07Product
OSuite
Reference implementation
Makes Ambit practical, and tests it under real workloads.
Cited in independent research. See the citations · Read the note
More research: Bounded Action Firewall · Agent Runtime Exposure Graph · Runtime Exposure Management
Papers are arXiv preprints. Matoi is a research release, not a hosted OSuite Studio rollout.
The protocol
Ambit: a common language for agent actions.
Ambit describes what an action may do, who may authorize it, within what bounds it runs and what proof remains. Its first layer, Action Effect Spec, and its test suite, Action Effect Bench, are public today and run without OSuite.
The protocol should not depend on OSuite. OSuite should depend on the protocol.
{
"spec_version": "0.1.0-draft.1",
"effects": [{
"category": ".../effects/funds-transfer/1",
"operation": ".../operations/refund/1",
"resource": { "state": "known", "value": "order-42" },
"quantities": [{ "value": "7300", "unit": "USD", "scale": "2" }],
"modality": "may"
}],
"coverage": { "scope": "direct", "completeness": "partial" },
"unknowns": [{ "code": "unobserved_descendants" }]
}$ node runtime-pilot/offline-run.mjs # Bench, offline
The product
OSuite makes Ambit practical.
OSuite is Ambit's reference implementation. It holds consequential agent actions for a person to approve, binds each approval to the exact action that runs and keeps replay-ready proof, so teams can use the protocol today and we can test it under real load.
$ curl -fsSL https://studio.osuite.ai/install.sh | bash How it works →
At Ond, we build governance infrastructure so AI agents can serve organizations without escaping human authority.
OSuite is Ond's first flagship product: the reference implementation of Ambit, built so teams can use the protocol today and so the protocol keeps meeting real systems.
Ond was founded by Jason Wang, who previously took Dify Japan from zero to $1M ARR. We have brought enterprise AI to market before; now we are building the layer underneath it.
10K+downloads
Mercarienterprise proof of concept
5papers on arXiv
4live listings: Claude, ChatGPT, n8n and Dify
Featured
