bounded-autonomy

Bounded-autonomy plugin harness for agents. Intent to implementation: ready, then go.

Give AI coding agents (Claude Code, Cursor, Aider, Codex) a structured autonomous loop with guardrails — boundaries, 5 verification gates, 3-layer self-reflection, and autonomous remediation. pip install ouro-loop. Zero dependencies.

GRAIL — A different approach to automation.

OCCP — OpenCloud Control Plane: Governance-first AI agent orchestration with Verified Autonomy Pipeline. Policy-gated, audit-logged, kill-switch protected. 14 MCP tools, 8 specialist agents, Telegram integration.

J-Stack overview and Harmony human-AI coexistence charter.

Proof of concept: a CPG trade-promotion deductions agent that adjudicates retailer claims with bounded autonomy — drafting (never executing) approve/deny/partial/escalate decisions with cited evidence, backed by an eval harness (ground-truth fixtures, graders, LLM judge, calibration gates) rigorous enough to trust on money.

A 29-hour field study of Ash operating Dwarf Fortress through bounded machine interfaces, and what happened when safety governance became backpressure.

Deterministic evaluation of AI agent runtime controls and bounded autonomy. Pinned systems, decision/effect checks, mutation testing, and explicit evidence limits. Includes ACS, revocation, and automotive assurance evaluations.

Research protocol and reference implementation for bounded action admission under explicit attention and risk constraints.

Reference lab comparing LLM workflows and bounded agents through reproducible multi-provider benchmarks, grounding, epistemic evaluation, and explicit authority boundaries.

Evidence-backed agent runtime controls for bounded tool execution: scoped capabilities, deterministic ALLOW/DENY/ASK decisions, tool-bound human approval, result gating, adversarial evaluation and authority incident evidence. Scoped ACS v0.1.0 reference implementation.

2026 AI Agent Loop v2.0: Autonomous Coding with Guardrails & Self-Healing