Infrastructure that lets agents find capabilities, act under clear rules and leave a verifiable record.
We work in three areas: agent infrastructure and orchestration, trust and governance for execution, and research into autonomous intelligence.
What autonomous systems need beyond intelligence.
An agent that acts on its own needs a way to reach capabilities, rules for what it may do, and a record of what it did.
AI agents now call tools, move data and trigger payments without a person approving each step. That makes the moment before an action, and the evidence after it, part of the engineering problem.
Our work covers three areas. Agent infrastructure and orchestration connects agents to tools and to other agents. Trust, governance and execution decides what may run and records what did. Autonomous intelligence research looks at what comes after.
Three complementary areas. Every capability carries the status its evidence supports.
How agents reach tools, data services and other agents.
A FeedOracle MCP server answered a live status call on 31 October 2026 and reported MCP and A2A support.
DemonstratedFeedOracle MCPEvidenceToolOracle publishes an A2A Agent Card (protocol 0.3.0) with a JSON-RPC endpoint and its skills. Checked on 31 October 2026.
DemonstratedToolOracleEvidenceRoutes an agent’s intent to suitable MCP and A2A endpoints, using an intent classifier with a static fallback route. The first handshake is free and stateless.
Implemented / documentedToolOracleA person stays in the loop for regulated or irreversible actions.
Implemented / documentedToolOracleA2A interfaces exist. Coordination of several agents on one task is not yet documented.
Status to confirm—Multi-step work carried out end to end by agents is not yet documented as a capability.
Status to confirm—What may run, and proof of what did.
Decides PERMIT or DENY for one exact action under the current intent and policy, before anything runs.
Implemented / documentedAgentNOMOSConfirms that the invoked capability is real and that its evidence is current and validly signed.
Implemented / documentedAgentNOMOSBinds execution to what was authorized and signs a receipt (Ed25519), checked against a separate execution JWKS. On 31 October 2026 we verified the published reference receipt independently: its digest recomputes, its signature verifies, and a changed field makes the check fail.
DemonstratedAgentNOMOSEvidencefeedoracle.io publishes a general JWKS (ES256K, ML-DSA-65) and a separate execution JWKS for AgentNOMOS receipts (Ed25519). Both checked on 31 October 2026.
DemonstratedFeedOracleEvidenceResearch directions of the FeedOracle Research Center, not finished results.
New methods, explored before anything is built on them.
In researchResearch CenterHow autonomous systems decide, act and stay accountable.
In researchResearch CenterGovernance for devices that lose their connection: offline authorization, locally signed receipts and queued recovery. An evaluation-stage internal prototype (EDGE-01), not tested on physical hardware.
In researchResearch CenterDetailsBeyond intent classification for routing, there is no documented capability yet.
Status to confirm—One concrete example from Trust, Governance & Execution: the sequence AgentNOMOS documents for every governed action.
The agent states what it wants to do.
A matching capability is found, for example through ToolOracle routing.
ToolOracleCurrent, source-bound evidence is attached to the decision.
AgentNOMOSPERMIT or DENY for this exact action.
AgentNOMOSOnly the authorized action runs.
AgentNOMOSA signed record of what actually ran.
AgentNOMOSThe observed result, recorded separately and linked to the receipt.
AgentNOMOSIndependent developments that grew out of this field.
Governs AI agents before they act and proves what happened after. Verification is stateless and works offline, with a published verifier and reference bundles.
A capability router for autonomous agents. It routes intents to MCP and A2A endpoints, publishes an A2A Agent Card, and keeps human approval for regulated or irreversible actions.
Where AI and autonomous systems meet the other three fields.
Agents pay per call with x402 in USDC on Base. Some paid endpoints are live; further routes are staged. A secondary connection: AgentNOMOS does not depend on a blockchain to authorize.
Implemented / documentedAuthorization rests on current, source-bound evidence. ToolOracle connects agents to data services.
Implemented / documentedQuantum feature maps for kernel methods: a compiled study whose documented IBM hardware run gave a negative result.
In researchThis field feeds the FeedOracle Research Center with open questions about how agents decide and stay accountable. Two of the five research directions start here. Work in them stays research until a prototype exists, and nothing is called demonstrated without a reproducible test.
FeedOracle Research CenterFor developers who want to build on this field, and for teams who want to research with us.
Start with the developments' own documentation and discovery files.
Bring a question about agent decisions, authorization or accountability to the Research Center.