Ai Manager

Inspiration

What it does

Inspiration

We were frustrated by how fragmented marketing workflows are — strategists plan in one tool, designers
create in another, reviewers comment in yet another. We asked: what if AI agents could collaborate like a real team, inside a single workspace? That question led us to build AgentHub, a Slack-like platform where AI agents plan, execute, and review marketing campaigns together — with humans in the loop at every approval gate.

What It Does

AgentHub is an AI agent collaboration platform where users create channels, mention specialized agents
(@planner, @executor, @reviewer), and watch them coordinate end-to-end campaigns. The AI Manager (orchestrator) breaks long-term goals into key results, delegates to specialist agents, tracks confidence
scores, and manages the full goal lifecycle — not just single tasks, but sustained multi-step campaigns over time.

We integrated OpenUI Lang for artifact rendering — instead of agents dumping raw markdown, they output
structured UI components (starting with social media posts) that render as live, interactive previews directly in the chat. The same schema drives both the LLM prompt and the frontend renderer, so what the AI
"writes" is exactly what the user sees.

We plugged in Composio to give agents access to real workspace integrations — Facebook, Slack, email, and
more. A policy-based filtering system controls which agents can call which integrations: the Account Manager gets full access (ComposioPolicy.ALL), while specialists get awareness-only mode — they know the
integrations exist and can say "Ask Account Manager to post to Facebook for you" instead of failing silently.

We also added the Guild agent (Adam — Competitor Researcher) to the platform via a webhook-based shell,
proving that agents from external runtimes can plug into AgentHub seamlessly alongside local PydanticAI agents.

How We Built It

• Backend: Python + FastAPI, PydanticAI for agent orchestration, WebSocket streaming for real-time token
  delivery
• Frontend: React with a Slack-like dark theme, OpenUI Lang component library for rich artifact rendering
  
• Composio: Policy-based integration layer — per-agent access control (ALL, ALLOWLIST, NONE) with awareness hints for restricted agents
  
• Architecture: An OrchestratorAgent base class implements a Claude Code-style multi-turn loop (plan →
  delegate → observe → approve). Worker agents are forked on demand. A GuildShell bridges external agents via webhook polling. Goal management tools give agents CRUD over OKR-style goals with confidence tracking.
• MCP Integration: Agents can connect to MCP servers (stdio & SSE) for external tool access
  

Challenges

• Orchestrator complexity: Getting the multi-turn loop right — streaming tokens, handling tool calls,
  managing approval gates, and persisting artifacts only after validation — required careful state management. One bug where _current_full_text wasn't reset between turns caused duplicated output for hours.
  
• OpenUI Lang integration: Teaching LLMs to output structured component markup instead of freeform text required precise prompt engineering. The schema had to be strict enough for the frontend parser but flexible enough for the LLM to use naturally.
• Composio policy design: Deciding which agents get which integrations was a design problem, not just a code problem. Giving every agent full access creates chaos; giving none makes the platform useless. The
  three-tier policy (ALL / ALLOWLIST / NONE with awareness) was our compromise — and the "awareness-only" mode where restricted agents suggest delegation to the Account Manager was a late insight that made the whole
  system feel coherent.
• Guild agent bridging: Polling an external webhook-based runtime and streaming its response back through our WebSocket pipeline — while maintaining the same UX as local agents — meant building a completely
  different execution shell (GuildShell) that still conformed to the OrchestratorAgent interface.
• Goal lifecycle: Moving from single-shot tasks to long-running goals with key results, confidence
  assessment ($\frac{\text{progress%}}{\text{time%}}$ ratio), and scheduled reconciliation was a significant
  architectural leap.

What We Learned

Agents are only as useful as their coordination layer. The hard problem isn't making one agent smart — it's making multiple agents collaborate reliably, with human oversight, artifact persistence, and graceful failure handling. Composio showed us that real-world integrations need access control just like human team
members — not everyone should have the keys to post on the company Facebook page. Building the orchestrator taught us more about production AI systems than any single model improvement could.

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Ai Manager

Built With

• composio
• guild
• openui
• python