Idea Sourcer
Research-backed ideation agent for finding, expanding, validating, and ranking ideas. Tuned past obvious AI wrappers — biased toward agentic systems with a sharp wedge.
Composable agent swarms — ideation, product, engineering, growth, marketing — coordinated by one orchestrator toward a single goal: ship a company.
Agent Orchestrator already runs parallel coding agents in isolated git worktrees — spawning them per issue, routing CI failures and review comments back, opening PRs on its own. That's the bedrock we build on.
On top of it, we're building a factory layer. Instead of one agent per ticket, you install a swarm — a specialist team with its own memory, prompts, and pluggable slots. Swarms coordinate through the orchestrator toward a single objective: shipping a company.
Ideate until one idea beats the novelty bar. Spec the product. Ship the PRs. Run the launch. All of it, declarative, version-controlled, reproducible.
Each swarm ships as a versioned folder of prompts, tools, and pluggable slots. Drop it in agents/, list it in your fleet manifest, and the orchestrator does the rest.
Research-backed ideation agent for finding, expanding, validating, and ranking ideas. Tuned past obvious AI wrappers — biased toward agentic systems with a sharp wedge.
Turns a winning idea into a spec: core loop, flows, scope cuts, buildable v1. Hands the brief to engineering.
Builds on top of Agent Orchestrator's core loop — parallel worktrees, autonomous CI recovery, review-driven iteration. A full engineering team expressed as configuration.
Acquisition experiments, funnel instrumentation, and conversion iteration. Runs against live traffic and reports what moved.
Positioning, launch narrative, distribution. Writes for the idea, not the template — and lines up the channels to ship into.
The fleet is a plugin system. Author a new swarm, drop it in agents/, and the orchestrator composes it into every run.
The orchestrator dispatches a goal across the fleet. Each swarm runs autonomously, emits artifacts into shared memory, and feeds status back. Nothing moves without the orchestrator knowing.
Every agent is a folder of prompts and a handful of pluggable slots. Swap the idea-generation slot from codex to another config, bump the novelty bar, pick a different research mode. No rewrites, no glue code.
A versioned directory with agent.json, prompts, and slots. If it's in git, it's reproducible.
Named extension points — like idea-generation — with JSON schemas you can't silently violate.
Drop a new config into a slot. No TypeScript, no bundler. The orchestrator resolves the graph at fleet-up.
{ "id": "idea-sourcer", "name": "Idea Sourcer", "description": "Research-backed ideation agent for finding, expanding, validating, and ranking ideas.", "slots": { "idea-generation": { "path": "./slots/idea-generation", "config": "codex" } } } // slots/idea-generation/configs/codex.json { "id": "codex", "researchMode": "deep", "noveltyBar": "high", "outputMode": "ranked-shortlist", "constraints": [ "Prefer ideas that feel impossible without Codex.", "Bias toward agentic systems that operate software.", "Avoid thin wrappers unless the wedge is sharp." ] }
Products the factory is building during the hackathon. Full details, live demos, and source on the projects page.
→ github.com/i-trytoohard/codex-startup-factory