Orchestrating Four AI Agents: Building orcha and council-cli
Every LLM fails differently. orcha explores what happens when Codex, Claude, Gemini, and Kimi work in parallel and are forced into an explicit synthesis loop.
orcha and council-cli are attempts to make model disagreement useful. Instead of hiding parallel agent work behind a single answer, the systems expose roles, critique, synthesis, and decision flow.
The throughline is orchestration as product infrastructure: assigning work to different model strengths, preserving shared context, and forcing the final result through a legible review ritual.
The problem with one-agent interfaces
A single model answer can be fluent and still miss important failure modes. The point of orcha is to make parallel model work structured enough that disagreement becomes useful rather than noisy.
That requires explicit roles, task boundaries, shared context, and a synthesis step that does not blindly average every response.
What orcha and council-cli explore
orcha focuses on dispatch and coordination across Codex, Claude, Gemini, and Kimi. council-cli focuses on turning model disagreement into a visible review process for planning and architecture.
The systems share a product thesis: multi-agent workflows need operational UX, not just more model calls.
Design lesson
The hard part is not making several agents talk. The hard part is making their work inspectable, bounded, and synthesizable so the human operator can trust the final decision path.
Related work
orcha
A multi-agent orchestration layer that distributes work across Codex, Claude, Gemini, and Kimi so parallel model effort becomes a usable execution workflow.
Core challenge: Making heterogeneous model strengths additive, structured, and operational instead of noisy.
- Parallel agent dispatch
- Cross-model synthesis
- Dependency-aware task routing
council-cli
A multi-agent coding roundtable that turns model disagreement into a visible review process for architecture, implementation planning, and technical critique.
Core challenge: Preserving shared context while letting specialist agents critique the same problem from meaningfully different angles.
- Shared context handoff
- Multi-agent review ritual
- Architecture-oriented critique loop