Search⌘ K
AI Features
Home
Courses
Google FDE Interview Prep

Agent-to-Agent Communication with A2A

Agent-to-agent communication facilitates task delegation across agent boundaries, allowing remote participants to manage evolving tasks. It emphasizes the importance of maintaining responsibility and validating assessments through authoritative records. The process involves various stages, including submission, progress updates, clarifications, and completion, with strict rules governing when to release assessments. An Agent Card provides essential metadata for selecting reviewers, but does not guarantee the quality of responses. The interaction must be carefully managed to ensure that task outcomes and business assessments are accurately represented, separating transport observations from final verdicts.

We'll cover the following...

The remote supplier-review agent reports, "Evidence collection has started," and then requests clarification: "Which review date applies to S-41?"

The coordinator has received a progress update and a request for clarification, not a completed assessment. It must not issue a readiness verdict before receiving and validating the required evidence.

The Agent2Agent (A2A) protocol allows a coordinator to delegate work to an independent agent that manages its own task execution and reports progress. The coordinator must track the remote task's state and handle follow-up requests rather than treating the first message as a completed tool result.

What owns the review?

The previous lesson restricted status retrieval to a single tool operation. Here, get_review_record retrieves a record, while the readiness-review agent performs the assessment. It retrieves two authoritative records, requests missing information, applies the review policy, and returns an evidence-backed assessment.

MCP standardizes access to tools and context sources, while A2A supports communication and task delegation between independent agents. Neither protocol determines an implementation's complexity or suitability for a particular task. A2A interactions can be brief or involve multiple exchanges. Delegation is appropriate when an independent agent is responsible for completing a defined task, not simply because the task is long-running.

For S-41, the review contract requires:

  • Insurance: Verify that the supplier's insurance coverage is valid on the specified review date.

  • Security: Confirm that the authoritative security-review record shows a passing result.

  • Output: Return a readiness assessment that includes the insurance and security findings, supporting evidence, and applicable review date.

  • Authority: The reviewer may assess supplier readiness but must not approve or initiate a purchase.

Inspect the Agent Card

An Agent Card advertises identity, description, and input/output capabilities. Inspect its purpose, skills, input/output modes, endpoint, and supported capabilities before selecting the reviewer. This is discovery metadata: It supplies no proof of answer quality, universal reachability, or downstream record authority.

Separate the objects involved before writing coordinator logic:

  • Message: One contribution to an interaction, such as the request or clarification reply.

  • Task: A tracked unit of delegated work with an execution outcome.

  • Context: A conversation scope that can associate related messages; retain it separately from the task identifier.

  • Status: The current execution condition, such as awaiting input or having finished.

  • Artifact: A produced output, such as assessment A-41, whose completeness and business meaning need validation.

A card describing “supplier readiness review” helps select a candidate participant. Acceptance still requires evidence that the implementation reads both sources and produces the promised assessment.

T-41 completes with assessment A-41: not ready, because insurance expired before the review date
T-41 completes with assessment A-41: not ready, because insurance expired before the review date

Gate release on completed execution and a validated assessment, not on ready=true. Otherwise, the coordinator would suppress valid negative assessments alongside incomplete work.

Follow T-41 completely

The labels below are application-level event categories, not A2A protocol state names.

Task T-41 belongs to context C-7. The insurance record expires on 2026-10-20; the security review is passed. The review date is initially absent.

  1. Submission: Send the S-41 assessment request. Display “Review requested.” Retain T-41, C-7, supplier identity, and the request. Release no verdict.

  2. Progress: Receive “Evidence collection has started.” Display it as progress. Retain correlation and the observation; do not interpret the text as an assessment.

  3. Clarification: Receive “Which review date applies?” Display that question. Retain the pending field review_date against T-41/C-7; pause the success path.

  4. Reply: Send 2026-10-27 correlated to the intended task and context using the selected interface’s continuation mechanism. Display “Date supplied.” Retain the reply and clear the pending question only when the application accepts that correlation. No verdict yet.

  5. Resumption: Receive renewed work progress. Display “Review resumed.” Retain the supplied date and current execution observations. Release no verdict.

  6. Completion: Receive assessment A-41, identifying S-41, review date 2026-10-27, insurance expiry 2026-10-20, security passed, ...