Principle 01
Three agents only
Generator, Evaluator, Combiner. Every benchmark gain has come from this trio — not from specialized roles. The simpler the colony, the easier it is to understand why it wins.
// PRINCIPLES
The architecture is frozen.
Five rules that keep Colony ant-like. Next insights come from benchmark data — not more design.
The Frozen Five
Principle 02
Three pheromones only
Novelty pulls outward. Evidence pulls toward reality. Constraint Fit pulls toward the question. Three dimensions create useful tension; eight become impossible to reason about.
Principle 03
Local context only
No agent sees the full notebook. Each one gets a sampled local view: strong trails, random trails, recent trails. Universal visibility kills emergence — agents stop being ants.
Principle 04
Exploration budget 70 / 20 / 10
Every local view reserves 70% attention for the strongest trails, 20% for the middle, and 10% for weak/random trails — plus a guaranteed recency slot. Real colonies never stop exploring.
Principle 05
Recombination is king
The strongest answers are hybrids: Idea A + Idea F + Idea M → new organizing principle. If one operation drives most colony wins, it's Combine — not generation, not scoring.
What we are NOT building
Every iteration that drifts toward a conventional multi-agent framework gets pulled back here. The temptations to refuse:
- ✕No planners
- ✕No reviewers
- ✕No red teams
- ✕No governance layer
- ✕No executive layer
- ✕No domain routers
- ✕No extra scoring dimensions
- ✕No agent that sees everything
Deferred experiments
Lineage-trace synthesis
Today the synthesizer picks the best note and writes the answer. Eventually we want it to trace the strongest lineage — Note 4 → Note 11 → Note 27 → Note 44 — and synthesize from the winning trail, not just the winning snapshot. That is more ant-like, but it is a real change to the final step. Post-benchmark candidate, not v1.
External access · MCP
Colony as a /think_harder endpoint
The colony is exposed as a single MCP tool so any agent (Claude Desktop, Cursor, other apps) can call it when it needs deeper reasoning. No notes API, no run browser, no streaming — one tool, one answer.
POST /api/mcp
Authorization: Bearer <MCP_SECRET>
tool: run_colony({
problem: string,
mode?: "framework" | "alternatives" | "creative" | "research", // default "framework"
max_ticks?: number // 1–12, default 6
})
→ {
run_id, answer, framework,
top_notes[], alternatives[],
score, stats { ticks, llm_calls, latency_ms }
}Internals stay internal. Ants don't expose their pheromone molecules — they expose "here's the food."
North Star
Can useful intelligence emerge from many simple agents interacting through a shared trail system?