aicl-protocol

AICL: Agent Intent Communication Language

AICL is a structured communication language for AI agents and human-agent collaboration. It prioritizes low ambiguity, high efficiency, and auditable decision flow.

🎮 Try it now: AICL Playground

Status: Final v1.0.0 (2026-02-25)

Why AICL?

Modern AI agents need to communicate with each other—and with humans. But existing protocols are either:

• Too low-level (JSON, REST)
• Too verbose (XML)
• Not designed for agent autonomy

AICL fills this gap with intent-first, constraint-aware, and negotiation-native messaging.

Quick Start

# 1. Install
pip install aicl

# 2. Parse an AICL message
aicl parse examples/01_ask.aicl --pretty

# 3. Validate messages
aicl validate --strict examples/

# 4. Transpile natural language to AICL
aicl transpile "please verify policy compliance for release"

Or use the online playground: https://haocn-ops.github.io/aicl-protocol/

Core Goals

• Intent-first messaging
• Constraint-bound execution
• Evidence-aware reasoning
• Negotiation-native collaboration
• Human-in-the-loop (HITL) safety controls

Message Model

MSG{
A:<Actors>
I:<Intent>
O:<Object>
T:<TaskGraph>
G:<Delegation>
C:<Constraints>
K:<Known>
U:<Unknown>
P:<Proposal>
R:<Risk>
Q:<Query>
D:<Decision>
V:<Votes>
M:<MemoryCapsule>
H:<HITL>
X:<Expectation>
S:<State>
}

Required Fields

• Always required: I, O, S
• Multi-agent context: A
• Delegation: T, G
• Negotiation/decision flow: D
• High-risk operations: R, H

Standard Intents (v1.0)

ASK, INFORM, PLAN, ACT, VERIFY, NEGOTIATE, DELEGATE, ACCEPT, REJECT, ESCALATE, SUMMARIZE, COMPARE, ESTIMATE, PRIORITIZE, SCHEDULE, BLOCK, UNBLOCK, REPLAN, CANCEL, PAUSE, RESUME, COMMIT, ROLLBACK, AUDIT, TRACE, CITE, CLARIFY, CONFIRM, DISPUTE, RESOLVE, SPLIT, MERGE, ROUTE, HANDOFF, MONITOR, ALERT, REPORT, DIAGNOSE, FIX, TEST, VALIDATE, SANITIZE, FILTER, TRANSFORM, RETRIEVE, SYNTHESIZE, CRITIQUE, JUSTIFY, PREDICT, CLOSE

Conflict Resolution Priority

policy > hitl > authority > evidence > feasibility > utility > vote

Minimal Validation Rules

• S.conf must be within 0.00..1.00
• S.ver uses semantic versioning (e.g. 1.0)
• S.trace must be globally unique
• No cyclic dependencies in T.deps
• H.required=true blocks high-risk COMMIT before human response

Example

MSG{
A:{from=agent_planner,to=[agent_risk],role=planner}
I:NEGOTIATE
O:release/2026Q1
C:deadline=2026-02-26T12:00+08;policy=P1
K:[data_ready,legal_partial]
U:[final_legal_clause]
P:[ship_with_guard_clause]
R:[compliance_risk_if_unguarded]
H:{required=true,trigger=policy_risk,question="Allow guarded release?",options=[allow,delay,reject],sla=2026-02-25T21:30+08,default_action=delay}
X:need_human_choice_id
S:conf=0.61;ver=1.0;trace=trc_aicl_001
}

Compare with Other Protocols

Feature	AICL	OpenAI Agents SDK	MCP (Anthropic)	REST/JSON
Intent-driven	✅	❌	❌	❌
Built-in Negotiation	✅	❌	❌	❌
HITL Safety Controls	✅	Partial	❌	❌
Constraint Propagation	✅	❌	❌	❌
Confidence Tracking	✅	❌	❌	❌
Human-Readable	✅	✅	✅	❌
Extensible	✅	✅	✅	✅

AICL is designed specifically for multi-agent collaboration with built-in support for:

• Task delegation with capability matching
• Risk-aware decision making
• Human-in-the-loop approvals
• Auditable trace logs

Roadmap

• Next: richer semantics and interoperability profiles

Repository Structure

• README.md: quick overview
• SPEC.md: full v1.0 draft specification
• ABNF.md: textual grammar draft
• examples/: ready-to-use message samples
• tools/validate_aicl.py: validator (supports --strict)
• tools/parse_aicl.py: AICL text to JSON parser
• tools/transpile_nl_to_aicl.py: NL to AICL draft transpiler
• tools/aicl_cli.py: unified CLI (parse, validate, transpile)
• tests/: unittest suite
• FINAL_VERSION.md: release summary

Validation

Run the validator against examples:

python3 tools/validate_aicl.py examples
python3 tools/validate_aicl.py --strict examples

Parse one AICL message into JSON:

python3 tools/parse_aicl.py --pretty examples/02_negotiate_hitl.aicl

Transpile a natural language request to AICL draft:

python3 tools/transpile_nl_to_aicl.py "please summarize weekly incidents"

Use unified CLI:

python3 tools/aicl_cli.py parse --pretty examples/02_negotiate_hitl.aicl
python3 tools/aicl_cli.py validate --strict examples
python3 tools/aicl_cli.py transpile "verify policy compliance for release"

Run tests:

python3 -m unittest discover -s tests -p 'test_*.py'

This site is open source. Improve this page.