Releases·Tool Release
CROA open-source project proposes deterministic execution layer that checks full agent trajectories, not just individual actions
Two developers released CROA, an open-source architecture and prototype Pilot that separates agent intelligence from execution authority. The external execution boundary evaluates whether an agent's proposed action keeps system state admissible before allowing state changes, catching constraint violations that individual action-level checks miss. The authors are asking agent-runtime builders to stress-test the design.
1 engagement1 sourceThu, Sep 17, 2026, 02:05 PM
AI Overview
- CROA separates agent intelligence from execution authority: agents propose actions while an external boundary verifies state admissibility before state changes
- The deterministic Pilot covers cumulative trajectory constraints, execution contracts, operation mutation, replay protection, and tamper-evident evidence
- CROA is open source and free with no commercial version or signup; code on GitHub and spec on Zenodo
- Authors cite trajectory risk from individually valid actions, e.g. three 40-record exposures crossing a 100-record customer limit
- Authors ask agent-runtime builders to test the architecture's execution paths and edge cases and report where the model breaks
Entities
Related
Products·launched
GraphARC: Governed agent runtime that checks and refuses actions
GraphARC is an open-source governed agent runtime where a model proposes a multi-node graph for a task, and a deterministic checker admits or refuses it with reasons before execution. It enforces policies, like denying rollback actions, to ensure agents only do what's allowed, solving the problem of unconstrained agent actions.
GraphARC is an open-source governed agent runtime where a model proposes a multi-node graph for a task, and a deterministic checker admits or refuses it with reasons before execution. It enforces policies, like denying rollback actions, to ensure agents only do what's allowed, solving the problem of unconstrained agent actions.
1 engagement1 sourceredditWed, Aug 5, 2026, 06:44 PM
Products·building
Deterministic pre-execution authorization boundary for AI agents
OxDeAI is an open-source protocol that enforces a fail-closed authorization boundary before agent actions execute. It uses signed artifacts and adapters for LangGraph, CrewAI, and AutoGen to prevent retry amplification, budget leaks, stale-state executions, and permission drift. It solves the problem of agents performing unauthorized or non-idempotent actions for developers building production agent systems.
OxDeAI is an open-source protocol that enforces a fail-closed authorization boundary before agent actions execute. It uses signed artifacts and adapters for LangGraph, CrewAI, and AutoGen to prevent retry amplification, budget leaks, stale-state executions, and permission drift. It solves the problem of agents performing unauthorized or non-idempotent actions for developers building production agent systems.
1 engagement1 sourceredditTue, Jul 21, 2026, 12:34 PM
Products·building
SOPVM: Runtime enforcing SOPs as executable authorization boundaries for LLM agents
SOPVM is an open-source runtime that treats Standard Operating Procedures (SOPs) as executable specifications, placing the authorization decision outside the LLM. It solves the problem of LLM agents performing irreversible actions without proper oversight, providing a boundary for what the model is allowed to do.
SOPVM is an open-source runtime that treats Standard Operating Procedures (SOPs) as executable specifications, placing the authorization decision outside the LLM. It solves the problem of LLM agents performing irreversible actions without proper oversight, providing a boundary for what the model is allowed to do.
2 engagement1 sourceredditTue, Aug 18, 2026, 07:55 PM
Products·launched
ARK — open-source runtime supervision layer that validates agent tool calls before execution
ARK is an open-source runtime layer that checks an agent's proposed tool actions against applicable constraints and trusted evidence before they execute. Agents propose an action, ARK returns ALLOW, REJECT, or REQUIRE_EVIDENCE, with feedback sent back to the agent. It targets developers building tool-using agents whose actions are consequential, such as refunds, bookings, or record modifications.
ARK is an open-source runtime layer that checks an agent's proposed tool actions against applicable constraints and trusted evidence before they execute. Agents propose an action, ARK returns ALLOW, REJECT, or REQUIRE_EVIDENCE, with feedback sent back to the agent. It targets developers building tool-using agents whose actions are consequential, such as refunds, bookings, or record modifications.
19 engagement1 sourceredditMon, Aug 31, 2026, 01:24 PM
Products·building
OIP — fail-closed authorization gate that blocks AI workflow actions missing evidence or approval
OIP (Organic Intelligence Protocol) is a prototype gate that AI workflows must pass before taking actions: it verifies that required evidence and authorization are present, and stops if a condition is missing. LLM-driven workflows call the gate so agents cannot execute unauthorized actions. It targets the problem of AI agents acting without approval, for developers building agent systems.
OIP (Organic Intelligence Protocol) is a prototype gate that AI workflows must pass before taking actions: it verifies that required evidence and authorization are present, and stops if a condition is missing. LLM-driven workflows call the gate so agents cannot execute unauthorized actions. It targets the problem of AI agents acting without approval, for developers building agent systems.
11 engagement1 sourceredditTue, Oct 13, 2026, 03:29 PM