---
title: 'AgentHub: operating and coordinating a fleet of Codex agents'
date: '2026-10-13T15:54:13+08:00'
summary: A product-change walkthrough shows how AgentHub combines fleet management, authorised agent-to-agent delegation and sequential workflows around the same deployed agents.
lang: en
category: AI/Agents
social:
- channel: linkedin
  url: https://www.linkedin.com/feed/update/urn:li:activity:7515682888621907969/
  published_at: '2026-10-13'
  article_hash: 0aadd99ba27c134c38b8c10cfd559456576249df91d5e45bcd45fe6a916b84af
---

A product team wants to turn customer feedback into a change it can release. Someone needs to clarify the requirements, implement the change, and review the result against the acceptance criteria.

With several coding agents, the team also needs to organise that work: choose the agents, pass useful results between them, see where execution stopped, and inspect what was delivered.

I am building **AgentHub**, an enterprise AI agent fleet control plane under Tenon, around two connected responsibilities: **operate the agents and coordinate their work**. Codex provides execution; AgentHub connects deployment and lifecycle management with agent-to-agent collaboration and workflow orchestration.

## Put a project agent into service

Take **Build**, a project agent that helps maintain an application. It needs a repository, a responsible person, a Feishu entry point, model access and a running environment.

AgentHub carries creation through to a usable agent: create and publish the Feishu application, associate its owner, provision a separate model-gateway key, and start the Codex runtime with a persistent project workspace. The agent becomes ready only after its Feishu connection is confirmed.

The fleet view keeps Build and the other deployed agents visible. The operator can see which are working, idle or disabled, then open their operating details and recent recorded conversations. Team members interact with the agents through Feishu; the current deployment's control plane is run by the designated administrator.

![AgentHub's fleet console, with working, idle and disabled project and personal agents](https://tenon.asia/assets/products/agenthub-cluster-en-dark.webp)

## Let an agent ask another agent for help

Some collaboration is a specific request rather than a predefined process. While handling a direct task, Build might need another agent to clarify a product constraint before proceeding.

AgentHub's internal **agent-to-agent collaboration (A2A)** lets an agent list its authorised peers, submit a task and retrieve the result. The platform queues the request, waits for the target to be available, and records its execution and outcome.

The useful change is that an agent can obtain help from an existing collaborator through a tracked task. The operator does not have to relay every request and response manually. Access still depends on the agents' ownership or an explicit directional grant.

Here, A2A describes AgentHub's internal task-delegation interface. Interoperability with the external A2A protocol is not implemented.

## Coordinate a product change through a saved workflow

For work with a repeatable order, the operator can select those same deployed agents and save a sequence of steps. In the product-change demonstration:

| Step | Agent | Task and handoff |
| --- | --- | --- |
| 1 | Scout | Review the feedback and product material; define constraints and acceptance criteria. |
| 2 | Build | Implement the change; carry forward the acceptance criteria with a patch and validation material. |
| 3 | Prism | Review the implementation against the criteria; prepare findings and a release recommendation. |

Each step receives the original task, its own instructions, and the previous step's final handoff. In the workflow path, explicitly supplied files are validated and passed on as read-only copies. The agents keep their own workspaces.

Build's instructions in this walkthrough need to preserve Scout's acceptance criteria in that handoff, so Prism has the basis for its review. Later steps do not automatically receive every earlier result or the full conversation history.

The run view shows the steps, their states, handoff results, available files and links to the underlying conversations. The operator can inspect the delivered work before deciding whether to release it.

![AgentHub's product-change workflow, showing Scout, Build and Prism with their step instructions and handoffs](https://tenon.asia/assets/products/agenthub-workflow-en-dark.webp)

This reduces the work of coordinating each handoff while preserving a visible delivery chain. It does not establish that a patch is correct merely because all steps completed; the outcome still needs review.

The current workflows execute in sequence. A failed step stops the run, and an execution interrupted by a restart is marked for review rather than silently repeated. The platform rechecks agent availability and authorisation before proceeding.

## Stop the work without losing the workspace

After a workflow, Build and Prism remain the same deployed agents and can be used for later tasks. When a project pauses, **Disable** attempts to stop its agent's runtime and disable its model-gateway key while retaining the workspace.

If either operation fails, the control plane records an incomplete result that needs attention. Re-enabling the agent requires the runtime to confirm that it is online again.

Management and collaboration meet here: stopping an agent also makes it unavailable for new collaborative work. The workflow cannot treat a disabled agent as an available execution node.

![Build's details in AgentHub, showing its model, owner, operating status and lifecycle actions](https://tenon.asia/assets/products/agenthub-ownership-en-dark.webp)

## Why operate this stack yourself?

Feishu already offers [enterprise agent products with skills, connectors and management capabilities](https://www.feishu.cn/content/article/7677519271848610746). AgentHub's reason to exist is the operating choice it offers teams building around Codex.

Feishu provides a familiar place for the team to interact. The deployment operator controls the Codex runtime image, container resources, persistent workspace and model-gateway access. AgentHub makes the same deployed agents available for direct work, authorised delegation and saved workflows.

That is useful for a technical team that wants to run agents on infrastructure it manages and reuse them across both everyday requests and coordinated delivery. It also means taking responsibility for maintaining that infrastructure and reviewing what the agents produce.

The longer-term direction is to make a managed agent fleet reusable team infrastructure: agents that can take on individual work and be composed into larger processes. Internal delegation and sequential orchestration are the implemented starting point for that direction.

[Explore AgentHub and its interface](https://tenon.asia/products/agenthub/)

*The screenshots show AgentHub's actual interface with demonstration data, not customer activity. The current deployment uses single-host Docker. Direct A2A requests carry text; file handoffs belong to workflows. Workflows are sequential and do not provide branching, parallel execution or approval nodes. Recent conversation viewing covers seven days of available local user/assistant records, not all Feishu chats or every tool event. Arbitrary model/plugin configuration and a unified high-risk-action approval system are outside the capabilities described here.*