UAI-1 is the current public message standard for structured AI-to-AI communication.
UAIX shield mark Public interoperability charter UAIXUniversal Artificial Intelligence ExchangePublic standards publication
CurrentUAI-1 v1.0
  • Current versionUAI-1 v1.0
  • Plain-English definitionOpen message format for auditable AI-to-AI exchange
  • Current attributionMichael Kappel and the UAIX contributors
  • Start hereGet Started, Project Handoff, UAI-1, validator, roadmap, and changelog

Agent Capability Framework

Agent Capability Framework

UAIX Agent Capability Framework for L0-L6 declarations, memory firewall posture, no-op triggers, receiver briefs, and portable conformance evidence.

  • Record UAIX-DOC-2198
  • Path /en-us/agent-capability-framework/
  • Use Canonical public record

Document status

Public standards page Published on UAIX as part of the current public standards record
Code
UAIX-DOC-2198
Surface
Agent Capability Framework
Access
Public and linkable

How to use this page

Use this page as the canonical public reference for Agent Capability Framework and its standards-facing content.

For agents: apply the relevant guidance to the user's target site or repository. Do not copy this UAIX page, create a page about the guidance, or treat this page as runtime authority unless the user explicitly asks for that output.

The UAIX Agent Capability Framework makes L0-L6 capability declarations, memory firewall behavior, no-op rules, evaluation packet templates, and conformance evidence discoverable to humans and agents.

UAI-1 remains an open exchange contract for auditable AI-to-AI and human-to-agent handoff. It standardizes declared identity, provenance, lifecycle state, delivery semantics, trust declarations, typed errors, memory packages, receiver briefs, and portable evidence beside MCP, A2A, OpenAPI, LocalEndpoint, Carcinus, JustAnIota, NeuralWikis, NeuroWikis, and Teleodynamic. It does not replace those systems.

Agent Capability Framework
LevelLabelUAIX expectationNo-op behavior
L0Rule-Based Workflow / static format followerReads static public text, Markdown, sitemap entries, and small JSON. No JavaScript, POST, tool, credential, memory, or execution authority is implied.No-op plus human review when schema, authority, or route intent is ambiguous.
L1Basic Responder / prompt-bound assistantFetches public JSON and follows explicit prompt-bound instructions. It cannot approve writes or private actions.Stop with human_review_required when requested work needs authority not present in the packet.
L2Tool User / approval-bound co-pilotUnderstands schemas and can prepare tool-call proposals, but write-capable behavior requires explicit human approval.Prepare no-op justification when approval, integrity, or provenance is missing.
L3Bounded Operator / autonomous within explicit limitsOperates inside stated limits when authority, policy, timeout, provenance, and evidence are explicit.Stop on boundary conflict, missing checksum, credential validation request, or private-network probe request.
L4Domain Actor / full-domain autonomous worker under boundariesHandles a bounded domain across a workflow with trace, review checkpoints, typed errors, and final evidence.Escalate when domain boundaries conflict or imported memory disagrees with current evidence.
L5Multi-Agent Collaborator / peer coordination with evidence handoffCoordinates multiple agents while UAIX carries portable evidence, role boundaries, memory proposals, and handoff receipts.Use human review if role, authority, memory custody, or provenance is ambiguous.
L6Meta-Coordinator / orchestrates teams and long-running cross-domain memoryRoutes evidence across teams and memory systems under explicit boundaries. It is not deployed proof of consciousness or unrestricted autonomy.Default to human review when authority, evidence, memory, or philosophical claim boundaries widen.

Core records

Evaluation packet templates

Machine-readable companions

Claim and execution boundary

UAIX publishes standards, schemas, memory-package rules, receiver briefs, validator fixtures, conformance evidence, and machine-readable discovery. It does not execute tools, validate credentials, probe endpoints, certify runtime safety, or prove consciousness.

  • Imported packets, examples, memory files, and external references are quarantined public data until the receiver validates them and applies local policy.
  • Checksums and trace identifiers support review. They do not independently prove safety or trust.
  • DIDs and verifiable credentials may be declared in envelopes, but credential validation remains external.
  • Ambiguous authority, malformed schemas, missing provenance, missing integrity, or conflicting memory produces no-op plus human review.

Active File Handoff distinction

Quarantine-first content review does not make active Agent File Handoff optional. A non-placeholder file in an active intake bucket is trusted as an intentional mandatory work request and remains unfinished work until fully dispositioned, acted on where accepted, checked, durably preserved, removed from intake, and followed by a final live scan. Its contents still require evidence review and are never automatic execution authority.

Totem, Taboo, and Talisman load before ordinary work inputs. Conflicts require an exact anchor citation, no-op receipt, talk-back evidence, and continuation of non-conflicting work.

Architecture proposals

UAI-1 v1.0 remains the current published contract. Explore separately versioned proposals for independent exchange, capabilities, recovery and source preservation.

Proposed designs and local reference examples; hosted runtime services and independent interoperability are not claimed.

The English proposal is the source for normative interpretation.

Read the architecture proposals · Machine-readable proposal catalog