DocsLog in

The AI trust layer between agents: identity, agreements, safeguards, and outcome intelligence for A2A.

Commit 20744dd6

20744dd65349Merged

aryaamin <[email protected]> · 2 months ago · 40 files · +11 added · ~29 modified

Browse files at this commit

Description

Add adaptive AI assurance engine

Details

Commit
20744dd65349
Parent
89b58109a0e6

README.mdmodifiedView file

@@ -1,23 +1,27 @@
11 # OpenClasp
22
3-Trust and behavioural intelligence for AI agents.
3+AI assurance for agent-to-agent agreements.
44
5-OpenClasp is an assurance layer built on top of the
6-[Agent2Agent (A2A) Protocol](https://github.com/a2aproject/A2A). A2A gives agents a standard way to
7-communicate; OpenClasp adds the trust context needed before, during, and after that communication.
5+OpenClasp predicts whether an external agent will complete a specific agreement—and tells your agent
6+what safeguards to require. It runs on top of the
7+[Agent2Agent (A2A) Protocol](https://github.com/a2aproject/A2A): A2A handles communication;
8+OpenClasp adds assurance before, during, and after it.
89
910 ## What OpenClasp adds
1011
11-- Authenticated publishers, verified cloud runtimes, and shareable agent cards.
12-- Explicit agreements, safeguards, and versioned amendments between agents.
13-- Direct A2A sessions with scoped, short-lived credentials.
14-- Authenticated structured outcomes and private bilateral feedback.
15-- Task-specific behavioural history and contextual reliability intelligence.
16-- Deterministic policy checks, delegation, receipts, disputes, and audit history.
12+- Predicts success for the exact task, agent, and version.
13+- Asks one high-value question at a time to expose material risk.
14+- Recommends safeguards before either agent commits.
15+- Learns which questions, answers, and safeguards correlate with better outcomes.
16+- Records agreements, evidence, outcomes, and private bilateral feedback.
17+- Verifies publishers and cloud runtimes; provides shareable agent cards.
1718
1819 Conversation bodies stay directly between agent runtimes. OpenClasp does not receive, relay, or
1920 store them.
2021
22+Claude powers the AI layer when `ANTHROPIC_API_KEY` is configured. A conservative deterministic
23+fallback keeps assurance available during cold start or provider failure.
24+
2125 ## Run locally
2226
2327 Requires Node.js 24+ and Corepack.