Cohestra: One Coordinator, Eight Specialist Agents
Note: This blog is a personal notebook. I write it to record what I work on, for my own reference, not to publish it on social media or to promote myself. If a post is still useful to you, I am glad, but that is not why it exists.
This post was published on November 13, 2026. Software versions, product features, links, and tool availability may have changed since then. Always check official documentation for the latest information. Some screenshots show small details. Zoom in on the page to see them clearly.
An engineering task needs several roles: exploration, design, implementation, tests, security checks, and documentation. A single AI session does all of them in one long context. The context grows, the roles blur, and the result is hard to check.
Cohestra is my answer. It is a public agent plugin with one visible agent, the mb-Cohestra Coordinator. The coordinator works like the conductor of an orchestra: it plays no instrument, but it decides who plays and when. It defines the result, gives bounded tasks to hidden specialist agents, checks their evidence, and returns one result. The plugin is MIT licensed, and version 0.1.0 is a product preview.
The agents are general software engineering roles. The README examples use AL development for Business Central.
The agents
A specialist is a hidden agent with one engineering role. You cannot select it. Only the coordinator can call it.
| Agent | Role | Copilot tools |
|---|---|---|
| Coordinator | Defines the result, delegates, checks outputs, reports one result | read, search, agent |
| Context Analyst | Finds files, symbols, dependencies, tests, and repository rules | read, search |
| Software Architect | Designs boundaries, interfaces, data flow, and migration steps | read, search |
| Implementation Engineer | Makes focused code and configuration changes | read, search, edit, execute |
| Security Engineer | Assesses trust boundaries, access control, input, and secrets | read, search, execute |
| Test Engineer | Designs, writes, and runs focused tests | read, search, edit, execute |
| Quality Engineer | Checks correctness, compatibility, performance, and regression risk | read, search, execute |
| Documentation Engineer | Updates affected user, API, and developer documentation | read, search, edit |
| Pattern Writer | Produces repetitive edits from a verified local pattern | read, search, edit |
The coordinator has no edit tool and no execute tool. It delegates every change and every command. This keeps its context small and the roles separate.
How a task runs
You select the coordinator and describe one result. The coordinator then works in this order:
Here is a typical AL prompt from the README:
Add a Customer Rating field to the Customer table with a table extension. Show it on the Customer Card with a page extension. Use the object ID range from app.json and the name prefix of the other objects. Add an AL test and update the documentation.
The coordinator can split this prompt across the Implementation, Test, and Documentation Engineers. You still see one result.
The guard rails
- One visible agent. Each specialist sets
user-invocable: false, so the Copilot agent picker shows only the coordinator. The coordinator setsdisable-model-invocation: true, so no other agent can call it. - No nested delegation. No specialist has the
agenttool, and each one setsagents: []. A validator script fails when a specialist breaks this rule. - Least tools per role. Each agent gets only the tools that its role needs, as the table above shows.
- No hidden parts. The package has no hooks, MCP server, background service, or telemetry code. The validator fails on a hooks folder or an
mcp.jsonfile.
Clients and install
The package follows the Agent Plugins 1.0 standard. The Copilot agent files in com.github.copilot/agents/ are the source of truth. A sync script generates the Claude Code copies from them.
| Client | Support | Install |
|---|---|---|
| VS Code Copilot Chat | Yes | Enable chat.plugins.enabled, then run Chat: Install Plugin From Source |
| GitHub Copilot CLI | Yes | copilot plugin marketplace add mbaic/mb-cohestra-agent-plugin, then copilot plugin install cohestra@mb-cohestra-agent-plugin |
| GitHub Copilot app | Yes | Customize → Plugins, add the repository as a marketplace, then install cohestra |
| Claude Code | Compatibility layer | claude --plugin-dir ./mb-cohestra-agent-plugin from a local clone |
For the reasons behind one source for several clients, read Agent Plugins 1.0: One AL Plugin, Three Clients.
How it is tested
The repository has three levels of checks. A validator checks the manifests, the agent visibility, the delegation allowlist, and the Claude Code copies. Unit tests run with Python and need no network. Behavioral evals send real prompts to a real model and grade the result. An eval costs model usage, so the README shows how to set a cost ceiling.
Risks and limits
- The plugin can change files. Some specialists have edit and execute tools. The client asks you to approve each action that its rules protect.
- Your code enters the model context. Check the data policy of your organization before you use Cohestra on private code.
- No fixed token saving. Rules such as “search before you read” can reduce context use, but Cohestra promises no fixed saving.
- It does not replace the AL compiler. The agents read AL source as text. Build the app in VS Code to confirm a change.
Closing note
One coordinator with narrow specialists gives each role a small context and gives you one result to review. Version 0.1.0 is a preview, and I expect the agent prompts to change. The code is public. Read the agent files, criticize them, and copy the parts that fit your own team.