What developers build with A2AC.
A2AC gives developers a practical way to define squads, route work, recall memory, run approved tools, and return task receipts. The examples below show the shape of the system without forcing every customer into one model provider or one deployment path.
A service mesh for AI agents.
A traditional service mesh manages traffic, security, and observability between software services. A2AC applies that idea to AI agents: it manages task handoffs, memory recall, approved tools, routing rules, failure records, and receipts across agent squads.
The WiFi mesh analogy.
A WiFi mesh keeps devices connected across access points. A2AC keeps work connected across agents and systems. The goal is the same at a higher level: reliable routing, clean handoffs, visible failures, and one place to understand what happened.
Protocol specs.
The public contract is intentionally simple: submit a task, recall relevant memory, receive a result or error, and inspect the receipt. Specific connectors and deployment details can vary by customer environment.
POST /tasks creates a work record and routes it to an approved squad, worker, model, tool, or system.
POST /recall retrieves approved context for a task and returns a memory receipt hash.
GET /.well-known/agent.json describes supported agent capabilities and discovery metadata.
POST /api/v1/a2a accepts structured agent messages where enabled for a deployment.
Define who can work.
squad:
id: support_triage
memory_scope: customer_account
agents:
- id: triage_agent
role: classify_request
model: approved_fast_model
tools:
- ticket_lookup
- policy_search
routes:
- when: customer_request
send_to: triage_agent
receipts:
detail: metadata_with_redacted_summary
Send clear work.
POST /tasks
{
"task_type": "customer_request",
"squad": "support_triage",
"input": {
"summary": "Customer asks about policy status",
"source": "ticket_1042"
},
"controls": {
"human_review": "if_high_risk",
"max_cost": "0.50"
}
}
Receipt shape.
A receipt should show what happened without forcing sensitive raw content into every log. Customers can choose more or less detail based on policy, deployment, and compliance requirements.
{
"receipt_id": "rcpt_01",
"task_id": "task_1042",
"tenant_id": "customer_account",
"squad": "support_triage",
"route": {
"policy": "customer_request",
"agent": "triage_agent"
},
"memory": {
"scope": "customer_account",
"receipt_hash": "sha256:..."
},
"tools": [
{"name": "policy_search", "status": "ok"}
],
"result": {
"status": "completed",
"summary": "Request classified and routed"
},
"cost": {
"model": "approved_fast_model",
"latency_ms": 812
},
"integrity": {
"result_hash": "sha256:..."
}
}
Developers should be able to test a squad locally, submit a task, inspect the route, and see the receipt before deployment.
Task records and receipts can be exported to existing monitoring or security tools through logs, webhooks, or customer-approved sinks.
Squads can use different approved models for different work. A fast model can handle routine routing while a stronger model handles hard reasoning.
Chimera for live browser work.
When a task depends on the live browser, Chimera gives the agent a DevTools-based view of the current page. It is useful for development checks, user-authorized business software, console inspection, screenshots, and link or text extraction.
A2AC moves the critical state out of chat text and into task records that software can inspect.
A squad can route work across approved models, workers, tools, and private systems without changing the task format.
The receipt trail can be consumed by other systems, not only viewed inside A2AC.
Implementation questions engineers ask first.
Use leases, queues, and idempotency keys for shared resources so two agents do not mutate the same record, file, memory entry, or browser session at the same time.
Tools are exposed as scoped capabilities. The agent can request an action, but policy, route checks, and human review can block high-risk execution.
Receipts can include model, runner, token, latency, and budget metadata so teams can inspect what each task or squad consumed.
Receipt events can flow to customer-approved logging, monitoring, data, or security tools instead of living only inside A2AC.
Why this is different from one provider's agent tool.
Native agent tools usually work best inside their own model and cloud ecosystem. A2AC is meant to coordinate across models, tools, business systems, browser runners, hosted workers, and private deployments while keeping one task and receipt format.