AACPv1.4
[ Architecture ]

Where AACP fits in your stack

An honest answer to the question every architect asks.

Not every agent system needs this.

If your agents already exchange structured JSON via well-designed internal APIs, you probably don't need AACP today. Those systems have already solved the coordination format problem.

AACP addresses the systems that haven't: agent frameworks built on top of LLMs where coordination messages are written in natural language by default, vary on every run, and are expensive and difficult to audit.

The question to ask about your own system:

"When one agent needs to tell another agent what to do, what format does that instruction take? If the answer is 'whatever the LLM writes' — read on."

AACP fills one specific layer

Task execution

LLM reasoning, document processing, decisions. This is where model quality matters most.

NOT AACP
Coordination content

What one agent says to another. Typed packets. Validated. Deterministic. Currently natural language by default in most agent frameworks.

← AACP
Routing and delivery

A2A handles agent task routing. MCP handles agent-to-tool access. AACP operates inside these message payloads.

NOT AACP

MCP, A2A, gRPC, Kafka — all excellent at what they do. They solve routing and delivery. None specifies what goes inside the coordination message itself. That gap is what AACP addresses.

Four architectures. One honest assessment.

Where AACP makes a difference and where it doesn't.

ENTERPRISE · NO LLM AGENTS

Mature enterprise systems

Services communicate via internal APIs, message queues, and structured JSON. No LLM in the coordination path. Well-designed systems that have already solved this.

Service A
JSON / REST
Service B
structured event
Service C
AACP fitLOW

You already have structured coordination. AACP adds little here unless you are introducing LLM agents into this stack — see Quadrant iii.

LLM AGENT FRAMEWORKS · NO AACP

LangChain, CrewAI, AutoGen today

Agents coordinate in natural language by default. Every coordination hop passes through an LLM. Messages vary on every run. Expensive, non-deterministic, and hard to audit reliably.

Orchestrator
"Please ask the HR agent to retrieve..."
LLM call
"HR Agent, could you fetch the employee..."
HR Agent
"Thank you. Finance Agent, please calculate..."
LLM call
...
Finance Agent
AACP fitHIGH

This is AACP's primary target. Every coordination hop goes through an LLM call that could be a $0.00 typed packet. Natural language varies every run and cannot be reliably audited.

ENTERPRISE · ADDING LLM AGENTS

Enterprise systems introducing AI agents

A mature enterprise is adding LLM agents alongside existing services. The API layer works fine. The new LLM agent coordination layer is the gap.

Existing Service A
JSON / REST ✓
LLM Agent
"Please retrieve the employee records..." ✗
LLM Agent
JSON / REST ✓
Existing Service B
AACP fitHIGH

The API layer is already solved. The LLM agent coordination hop is where AACP applies. This is the most common enterprise entry point.

LLM AGENT FRAMEWORKS · WITH AACP

Agent frameworks with AACP coordination

LLM reasoning handles tasks that need it. Agent coordination is typed packets. Deterministic, auditable, cheap. The LLM focuses on the work, not the routing.

Orchestrator
FETCH|HR|return:ORCHESTRATOR|p:1|aacp:1.4|...
HR Agent ✓ validates. $0.00
MERGE|HR|return:ORCHESTRATOR|p:1|aacp:1.4|...
Finance Agent ✓ validates. $0.00
LOG|HR|return:AUD-Agent|p:2|aacp:1.4|...
Audit Agent ✓ deterministic. $0.00
AACP fitPRIMARY USE CASE

Every packet validated before transmission. Same packet every run. Machine-readable audit trail without post-processing.

Honest framing

AACP is most valuable when:

  • → Multiple LLM agents coordinate across several hops
  • → The same workflow runs repeatedly
  • → Auditability and compliance matter
  • → You need predictable, near-zero coordination cost

AACP adds less value when:

  • → Agents are doing open-ended research or creative work (task tokens dominate, coordination overhead is small)
  • → You have fewer than 2-3 agent hops per workflow
  • → Each workflow is genuinely novel and never repeats

Best-suited environments

Repetitive structured workflows

If the same workflow runs more than a few times a month, a rule-based encoder produces it at $0.00 indefinitely. Payroll, IT provisioning, invoice processing, contract review, sales qualification. 241 pre-built community rules cover the most common cases.

Compliance-sensitive pipelines

When you need to prove what instruction was sent to which agent, when, and in what form — typed packets are the answer. Every AACP packet is a machine-readable audit record without post-processing. The audit agent in the lab runs at $0.00 because it needs no LLM. IETF Internet-Draft draft-mackay-aacp-02 provides the formal specification for regulated environment adoption.

Multi-model agent systems

When agents run on different models from different vendors, natural language coordination creates subtle interpretation differences. AACP packets are model-agnostic by design — validated on Claude, GPT-4o, GPT-4.1, and GPT-4.1-mini. Same packet. Same behaviour.

Cost-predictable agent infrastructure

Known workflows: $0.00 encoding cost on every run. Novel patterns: $0.0020 once, $0.00 on every subsequent call. Amortisation benchmark: 91.6% cost saving across 240 encoding operations.

Honest framing on what AACP turned out to be most useful for: field notes from building the protocol.

Read the spec →See the lab results →Community rules (241) →