Deterministic coordination for multi-agent systems
AACP is an open protocol and encoder framework that transforms natural-language agent instructions into typed, auditable coordination packets. Same input. Same output. Every time.
"Please retrieve the employee salary records for the period ending 31 August 2024. I need all active employees, their departments, cost centres, base salary, any changes made this month, and pension contribution rates. Return as JSON array."
FETCH|HR|return:HR-Agent|p:1|aacp:1.4 |res:emp_salary|period:2024-08 |filter:status=active|fmt:json
Token reduction (~23% measured) is a secondary benefit. The primary value is determinism and auditability.
02 · Architecture
How AACP works
For novel workflows without a pre-built encoder, a three-tier fallback routes the instruction to an LLM once, logs the result, and serves it from cache on every subsequent call. One LLM encoding call per novel pattern, reused indefinitely at $0.00.
03 · Comparison
Traditional vs AACP coordination
Same outcome: a monthly department report. Fundamentally different coordination.
Traditional coordination
Report generated, eventually.
AACP coordination
Report generated faster, auditable, identical every run.
Task reasoning stays fully active in both workflows. AACP coordinates; agents think.
04 · Principles
Why structured coordination?
Deterministic
The same workflow produces the same packet every time. Known workflows encoded at zero LLM cost, identical output on every run.
Auditable
Coordination messages become structured, machine-readable records. Every packet can be logged, replayed, and inspected without post-processing.
Reusable
Workflow encoders can be shared across teams, systems, and organisations. 241 pre-validated community rules available at launch.
Model-agnostic
Validated on Claude Sonnet 4.5, GPT-4o, GPT-4.1, and GPT-4.1-mini. The same packet, the same behaviour, regardless of which model receives it.
Efficient
Measured coordination token reduction of approximately 23% versus verbose English. Validated from live API usage_metadata, May 2026.
05 · Encoders
Workflow encoders
Rule-based encoders transform common business workflows into coordination packets with zero LLM calls and deterministic output. Eight encoders across six domains, each grounded in a documented real-world platform.
HMRC PAYE / UK payroll practice
Microsoft Entra ID provisioning
AP three-way match
NDA/MSA review
Salesforce Agentforce 2026
ConductorOne, Lumos, CloudEagle
Zendesk Resolution Platform
NetSuite Autonomous Close
from aacp.encoders.workflows.payroll import PayrollEncoder
enc = PayrollEncoder()
pkt = enc.fetch_employees("2026-03")
print(pkt.packet)
# FETCH|HR|return:HR-Agent|p:1|aacp:1.4|res:emp_salary|period:2026-03|filter:status=active|fmt:json
print(pkt.api_cost_usd)
# 0.006 · Community
Community rule library
241 pre-validated coordination rules across 7 domains: HR, Finance, IT, Sales, CS, Legal and Marketing. MIT licensed. Load locally. Zero API cost.
07 · Ecosystem
Built alongside existing agent standards
MCP (Anthropic / AAIF)
Tool access and external system integration. AACP structures the coordination payload that travels inside MCP messages.
A2A (Google / AAIF)
Agent discovery and task routing. AACP structures the content of messages that A2A routes between agents.
Existing APIs and JSON
Internal service-to-service communication. AACP complements structured systems by providing a shared coordination vocabulary for the LLM agent layer on top.
08 · Integrations
Framework integrations
Drop-in adapters replace LLM-driven coordination calls with deterministic AACP packets. Measured against a full department-day workflow (59 coordination calls).
Positioning
AACP is not a runtime, orchestration platform, or model framework.
It is a lightweight coordination protocol and encoder framework designed to make agent workflows more structured, deterministic, and observable.
09 · Blog
Latest from the blog
What auditability actually means in a multi-agent workflow
Auditable AI usually means logging model outputs. AACP treats structured coordination messages as the audit trail itself. The difference matters when an auditor asks for it.
Read article →
I tested AACP against four agent frameworks. Here is what I found.
Same workflow, same data, same model, four frameworks: LangChain, CrewAI, AutoGen and Pydantic AI. 59 coordination hops each. The saving scales with how verbose the framework's default coordination is.
Read article →