Skip to main content
Telara

A2A Protocol / Overview

A2A Protocol

In Development

Agent-to-Agent (A2A) is Google's open protocol for direct communication between AI agents — no human in the loop, no shared tool runtime. Telara will expose an A2A endpoint so any compatible agent can delegate tasks directly to your knowledge base.

A2A Protocol
Agent Cards
Task API
Push Notifications
This page describes the intended design

A2A support is under active development. The architecture described here reflects the planned implementation — details may change before general availability. This page exists to document design intent and gather feedback.

Why A2A

How A2A differs from MCP

MCP and A2A solve different problems. MCP is a tool-serving protocol — a human-driven client calls tools on a server. A2A is an agent-routing protocol — one autonomous agent delegates a full task to another. The two are complementary, not competing.

MCP (what exists today)
A human-driven client (Claude Code, Cursor) calls individual tools
Request/response per tool call — one tool at a time
The client orchestrates when and how tools are called
Tools are discoverable through the MCP handshake
A2A (coming soon)
An autonomous agent sends a full task — Telara figures out how to complete it
Streaming and async — long-running tasks with progress updates
Telara orchestrates internally — the caller just gets the result
Agent Cards declare capabilities — no protocol handshake needed
When to use each
MCPYou're building a tool-augmented AI in an IDE or agent framework that calls individual search/action tools on demand.
A2AYou're building a multi-agent system where one agent needs to delegate a complex knowledge task to Telara and get back a structured result.

Architecture

How A2A with Telara will work

Telara will expose an A2A-compliant endpoint. External agents discover it via the Agent Card, then submit tasks using the standard A2A Tasks API. Telara routes tasks through its internal MCP engine and streams results back.

External Agent
Caller
Discover
Agent Card
tasks/send
HTTP POST
Telara
Task Router
Execute
MCP Engine
Result
Task complete

Discovery

Agent Card

Every A2A agent publishes a JSON Agent Card at /.well-known/agent.json that declares its name, capabilities, and skills. External agents fetch this card to understand what Telara can do before sending tasks.

/.well-known/agent.json
{
"name": "Telara Agent",
"description": "AI agent with access to your engineering knowledge base",
"url": "https://a2a.telara.dev",
"version": "1.0.0",
"capabilities": {
"streaming": true,
"pushNotifications": true
},
"skills": [
{ "id": "search_knowledge", "name": "Search Knowledge Base" },
{ "id": "browse_context", "name": "Browse Connections" },
{ "id": "execute_action", "name": "Execute Permitted Actions" }
]
}

The Agent Card is auto-generated from your Telara configuration. Skills reflect the data sources and permission policies attached to the configuration used for A2A access.

Task API

How tasks work

Telara's A2A endpoint will implement the standard A2A Tasks API. An external agent sends a task with a natural language message. Telara processes it — searching, browsing, and reasoning — then returns the result as a structured message.

tasks/send

Submit a task and wait for the complete result. Synchronous — best for quick queries.

Returns when the task is complete
tasks/sendSubscribe

Submit a task and receive streaming updates via Server-Sent Events as Telara works.

Streams intermediate steps + final result
tasks/get

Poll the status and result of a previously submitted task by its task ID.

For async workflows that don't block
tasks/cancel

Cancel an in-progress task. Telara will stop processing and return a cancelled status.

Best-effort cancellation

Task Lifecycle

Task states

Every task moves through a defined set of states. External agents can track progress either by polling tasks/get or by subscribing to streaming updates.

submittedworkinginput-requiredorcompletedfailedcanceled
input-required state

If Telara encounters an action that requires human approval (based on your permission policy), the task enters input-required state and sends a push notification to the configured webhook. Once the action is approved in Telara, the task resumes automatically.

Async Updates

Push notifications

For long-running tasks, Telara will support push notifications — webhooks sent to a URL you specify when submitting the task. This allows fully async, fire-and-forget workflows without polling.

Task complete

Webhook fired when a task finishes with the final result

Input required

Webhook fired when approval is needed before proceeding

Task failed

Webhook fired if Telara encounters an unrecoverable error

Authentication

How A2A auth will work

A2A access will use the same API keys you already generate from your Telara configuration. The same key that powers your MCP connection will also authenticate A2A task requests — no separate credential system.

Request Auth
Header
Authorization: Bearer telara_mcp_...
Same key as your MCP connection
Scope

The API key determines which configuration is loaded — and therefore which data sources and permission policies apply to the A2A task. The same deployment-based access model from MCP applies to A2A tasks.

Under the Hood

How Telara processes A2A tasks

Telara's A2A layer is a routing and orchestration wrapper around the same MCP engine that powers IDE connections today. When a task arrives, Telara's agent planner decides which tools to call and in what order — the external agent never needs to know.

1
Task received

External agent POSTs to tasks/send with a natural language message and optional skill hint.

2
Agent planner

Telara's internal agent analyses the task and selects the right sequence of MCP tools — search, browse, read, or action.

3
Tool execution

Tools are called against the configuration's data sources. Permission policies apply to any write actions, triggering approval flows as configured.

4
Result construction

Results are assembled into a structured A2A message with parts (text, data, files) and returned as the task output.

5
Delivery

The final message is returned inline (tasks/send), streamed (tasks/sendSubscribe), or delivered via webhook (push notification).

Roadmap

What's planned

Phase 1In progress
Core task API
  • tasks/send
  • tasks/get
  • tasks/cancel
  • Agent Card endpoint
  • Bearer token auth
Phase 2Planned
Streaming + push
  • tasks/sendSubscribe (SSE)
  • Push notification webhooks
  • input-required state
  • Approval flow integration
Phase 3Planned
Enterprise
  • Agent-level API keys
  • Per-agent permission policies
  • Task audit log
  • Rate limiting
Talk to us

If you're building a multi-agent system and want to integrate with Telara over A2A, book a demo. We provision workspaces for design-partner teams after a short walkthrough.

Book a demo