DEV Community

Cover image for Agent Taskbay: A Human-in-the-Loop Console for A2A Agents
Shashi Kanth
Shashi Kanth

Posted on Originally published at allsrc.dev

Agent Taskbay: A Human-in-the-Loop Console for A2A Agents

Your AI agents are running. Who is watching when they need a human?

One agent is waiting for deployment approval. Another needs clarification before it can continue an investigation. A third has been running for hours, and the engineer who started it has closed the browser. The agents may all speak A2A. That shared protocol does not, by itself, give the team one place to see the pending decisions, assign an owner, or look up exactly what was approved.

Agent Taskbay is the answer to that operational gap. It is a durable, self-hosted human operations console for A2A agents. It does not take over their execution. The npm package is agent-taskbay (MIT), and version 1.0.0 is the first stable release. The cover image is Home: what needs a person, the recent conversations, and the agents already connected.

TL;DR

  • The gap. Many agents means many places a person can be asked to decide, and a closed browser does not hold that work.
  • What Taskbay is. One inbox and one record for A2A agents. The agents keep execution.
  • What a person can do. Approve, edit and then approve, reject, request changes, delegate, steer, or stop and start again. Claim work, assign it, and set a due time.
  • What makes an approval hold. The decision names one revision. A newer proposal makes the old decision a 409. Nothing is sent until someone reviews the current revision.
  • Try it. Node.js 22.19 or newer, then npx agent-taskbay.

Why it exists

A chat window can hold one conversation with one agent. A team running several agents needs somewhere else. Questions, approvals, and tasks that run for minutes or days have to survive a closed laptop, and the person who should answer is often not the person who started the run. A decision kept in that place has to show what was approved, who approved it, and that the same version was the one sent.

The same rule, inside a single harness, is the HITL approval gate on allsrc.dev: approve an action, not a session. Taskbay carries it onto A2A, so any agent that speaks the protocol can use the console, in whatever language or framework it is written.

Architecture: human operators work in Agent Taskbay, which talks A2A to agents for coding, operations and research.

Execution stays with the agents. Taskbay is the console in front of the protocol.

Why not just use my agent framework's HITL support?

If you use LangGraph, Google ADK, or another agent framework, you may already have human-in-the-loop capabilities. You can pause execution, request human input, and resume when a decision arrives.

Agent Taskbay does not replace those mechanisms. It solves a different problem.

Framework-level HITL is an execution capability. Taskbay is a human operations capability.

Framework HITL makes an agent interruptible. Agent Taskbay makes human intervention operationally manageable across A2A agents.

Consider a system with several agents:

  • A deployment agent is waiting for approval to release a new version.
  • An investigation agent needs clarification before continuing.
  • A research agent has completed a report that requires human review.
  • Another agent is still executing, but a human needs to provide additional instructions.

Each agent may use a different framework. Each framework can manage its own execution state, checkpoints, and interruptions.

But where does an operator go to see all these requests? Who owns each decision? What happens when the assigned reviewer is unavailable? How do teams track pending work, escalate overdue tasks, or inspect what happened afterward?

These are operational concerns, not execution concerns.

Agent Taskbay provides a shared human-facing console for A2A-compatible agents. It brings task visibility, human requests, approvals, assignments, and audit history into one place.

The division of responsibility is intentional:

  • Agent framework: owns execution, state, checkpoints, interruptions, and resumption.
  • A2A protocol: provides a standard interface for communication and task interaction between compatible systems.
  • Agent Taskbay: provides the operator experience for interacting with those tasks, making decisions, and tracking human work.

Your agents continue to run in their existing frameworks. Taskbay does not need to become their orchestrator.

The objective is not to introduce another way to implement HITL inside an agent. It is to make human participation manageable when multiple agents, tasks, and operators are involved.

A framework pause shows up in Taskbay when the agent puts the task in input-required and advertises the approval-request extension. Typed questions use the structured-form extension the same way. The agent, in its own runtime, checks the decision before it performs the action. Taskbay supplies the reviewed decision and the record of the send.

What the operator gets

Home counts reachable agents, items waiting on a person, active tasks, and unread notifications. Inbox is the list across agents: tasks that need input, and approvals still open. The views are what needs you, what is yours, what is overdue, what is active, and what is done. The list can be narrowed by risk.

An approval is one proposed action on one task. A reviewer with operate access settles it in one of five ways: approve, edit and then approve, reject, request changes, or delegate. Approve sends the reviewed action once, after a confirmation. An edit is recorded as a new revision and sent. Reject and request-changes send nothing. Delegate hands the open request to another eligible reviewer. Every outcome except a plain approval needs a rationale. Separation of duties is on by default, so the person who opened a request cannot decide it. An approval the agent opened has no requester. An assigned request is decided by the assignee or an administrator.

The README also covers the work around the decision. A person can claim a task, assign it to someone who can operate that agent, and set a due time. An administrator names who receives overdue work, for the organization or for one agent, and overdue tasks are escalated once to that person. Extra input to a running task is reported as delivered, answered, refused, or still unknown. Stop and restart with this cancels the task and starts a new one in the same context.

From a request to a recorded decision

Connect an agent by its Agent Card, start a task, and the inbox is where the human part lands.

Agents page listing three sample agents from the local demo, each on 127.0.0.1 port 4010.

Three demo agents, connected by card URL.

Inbox in Agent Taskbay with Needs you selected. A pending approval titled Delete the staging cluster is marked high risk, beside tasks that are waiting for input.

Inbox, on Needs you. A high-risk approval sits in the same queue as ordinary questions.

Task detail for a sample task that is waiting for input. History shows the approval request, and the task audit records that the agent asked for a decision.

A task page: history, and the audit line for the approval the agent requested.

The proposal is stored as a numbered revision and is not edited in place. Its digest is the SHA-256 of the canonical JSON of the action, with object keys sorted at every level. The reviewer names the revision they read. The field list, the media type, and the message that carries the digest back to the agent are in the approval-request extension.

Here is the rule with a deploy as the example.

  1. The agent proposes revision 1: deploy version 2.1 to staging.
  2. The reviewer opens the approval and sees revision 1.
  3. The proposal changes. Revision 2 is now deploy version 2.2.
  4. The reviewer approves revision 1. Taskbay answers 409. The stale decision authorizes nothing. The reviewer has to read the current proposal before anything is sent.

An edit by the reviewer works the same way. It becomes a new revision, with its own digest, and that revision is what gets approved. A rejection or a request for changes sends nothing. The decision and the command commit together, the approved action goes out once, and the page records whether the delivery succeeded. An unknown delivery is not sent again.

Administration audit trail in Agent Taskbay, with People, Teams, and Audit trail among the tabs. Entries include agent registration and the request to approve deleting the staging cluster.

The audit trail. Decisions, revisions, and the delivery sit in an append-only record.

Three roles cover the console: admin, operator, and viewer. Teams, and grants per agent and per skill, decide who can read and who can operate. The trail keeps the proposal, the digest, the rationale, the reviewer, the delivery, and the ownership events. Database triggers reject updates and deletes of decisions and revisions. How commands are stored before they are sent, and how the screens keep reading that record after a restart, is covered in durable tasks. Ownership, due times, and escalation are in human in the loop.

The project is on the Allsrc projects page. a2a-wrapper exposes Claude, Codex, Copilot, OpenCode, and Antigravity as A2A servers, each with a card Taskbay can connect to. Source is in the repository.

Try it

You need Node.js 22.19 or newer. No database server, no account, and no .env file.

npx agent-taskbay
Enter fullscreen mode Exit fullscreen mode

The console listens on http://127.0.0.1:3002, opens the browser, and signs you in as a local administrator. The database is ~/.agent-taskbay. The process binds to this machine only.

In a second terminal, start the sample agents. They are fixtures.

npx agent-taskbay demo-agent
Enter fullscreen mode Exit fullscreen mode

It prints five Agent Card URLs on port 4010. Use the showcase card. Open Agents, choose Add agent, and paste:

http://127.0.0.1:4010/showcase/card.json
Enter fullscreen mode Exit fullscreen mode

Choose Fetch Agent Card, then Continue, then Connect. Open the agent, choose Start chat, and send:

please approve the deploy
Enter fullscreen mode Exit fullscreen mode

Open Inbox. Under Needs you the pending approval is "Delete the staging cluster", marked High risk. Review the proposed action, choose Approve, and confirm. The approval becomes approved, the content you reviewed is sent once, and the task completes. Administration → Audit trail lists the steps.

The quickstart covers the same path from a checkout, including a task started elsewhere. Issues go to the repository.


Originally published on allsrc.dev.

Top comments (0)