Conductor: A local multi-agent AI orchestrator that builds its own team

Technical automation

Everyone is demoing AI agents. Nobody explains how to build one. Conductor is a local orchestrator that routes tasks to specialist agents, creates new ones when none exist, and synthesizes the results — all from a single browser session on your own API key.

June 3, 2026 · By Sagheer Ahmed

I watched videos about AI agents for weeks. Everyone excited. Everyone showing demos. Not one video explained how to actually build one yourself — how to route between agents, how to create specialists on demand, how to get a coherent result from multiple models talking to the same problem. Just demos.

So I built it. Conductor is a local multi-agent orchestrator that runs in your browser, on your own Anthropic API key, on your own machine. Nothing goes to a third-party service. No subscription. No walled garden.

The source is on GitHub.

What it does

You describe a task in plain language. Conductor reads its library of specialist agents, picks the right ones for the job, runs them in sequence, and synthesizes a single coherent answer from all their outputs. It also shows you every action in a live panel on the right side of the screen as it runs.

The part that makes it interesting: if no agent exists for what you need, Conductor creates one from scratch, saves it permanently, and uses it immediately — all in the same session.

The demonstration that made it real

The best way to see what this means is through the religion demo.

I gave Conductor one prompt:

> "I want to do a comparative study of world religions — Christianity, Islam, Judaism, Hinduism, and Buddhism. Compare how each answers the question: what happens after death?"

There were no religion specialists in the library. Conductor:

Recognized the gap and decided to create five permanent specialists
Built ChristianityAgent, IslamAgent, JudaismAgent, HinduismAgent, BuddhismAgent — one by one, each with domain-specific instructions, boundaries, and example use cases
Called each specialist with a focused question about the afterlife, capturing internal variations (Sunni vs. Shia, Theravada vs. Mahayana, Catholic vs. Orthodox)
Synthesized a graduate-level side-by-side comparison across all five traditions
Scaffolded a ComparativeReligionStudy project folder to hold future work

All from one prompt. The five agents are permanently saved and available to every future session.

That combination — create, use immediately, keep forever — is what makes agent-on-demand different from just calling an LLM with a system prompt.

How it works

There are four moving parts.

AgentLibrary: A folder of YAML files. Each file is one specialist — a name, a domain description, a system prompt, clear boundaries (what it does and does not handle), example use cases, and tags. Conductor reads these at the start of every session to know what it has available.

Conductor (Opus 4): The routing brain. Given your task, it reads the library, decides which agents to call, in what order, with what context. When no agent fits, it calls create_agent to draft and save a new YAML definition before proceeding.

Specialists (Sonnet 4.6): Single-turn domain calls. Each specialist gets a self-contained prepared task from Conductor, runs once, returns its output. No memory between calls — Conductor carries the context forward.

Session and SSE streaming: A background thread runs the agentic loop. The browser receives server-sent events for every tool call — routing decisions, agent creations, specialist completions — and renders them in the activity panel in real time.

The model split is deliberate. Opus 4 for Conductor keeps routing decisions and agent creation at full quality. Sonnet 4.6 for specialists keeps the per-task cost low. Prompt caching on system prompts reduces repeated token costs as a session grows.

A typical multi-agent task — three to five specialists — costs between $0.20 and $0.80 on the Full preset. The Economy preset (Sonnet for all roles) brings that down to $0.05 to $0.20.

Running it

BASH

git clone https://github.com/SagheerDBA/Conductor.git
cd Conductor
pip install -r requirements.txt

Set your API key:

BASH

# macOS / Linux
export ANTHROPIC_API_KEY='your-key-here'

# Windows PowerShell
$env:ANTHROPIC_API_KEY = 'your-key-here'

Then:

BASH

python app.py

Open http://localhost:5002. Select Full or Economy, click Start Session, describe your task.

Adding your own agents

Two ways.

The easier way: describe a task that needs a specialist. Conductor will draft the full YAML definition, ask if you want to save it, and once confirmed, use it immediately. Next session, it is just there.

The manual way: copy the schema from AgentLibrary/, fill in the fields, and save to:

AgentLibrary/Work/agents/<slug>/definition.yaml for professional or technical domains
AgentLibrary/Personal/agents/<slug>/definition.yaml for personal domains

What this is not

Conductor is not a production orchestration framework. It does not have retry logic or rate-limit handling. It is a working tool for personal and professional use — good enough to get real work done, simple enough to read and modify.

If you want something to route complex tasks across domain specialists without building a framework from scratch, this is a reasonable starting point.

The code is at github.com/SagheerDBA/Conductor.