Agentic Fusion

Agentic Fusion is Clioloop's way of producing frontier intelligence on open models. Instead of asking one model to do everything, Fusion surrounds your main model with a panel of helpers: planners propose routes, your model does the visible full-tool work, independent reviewers critique the draft, and it all fuses into one answer. The quality comes from the synthesis — not from running the same prompt five times.

Why fuse models?

A single cheap or open-weight model has blind spots. Fusion turns those blind spots into a strength by giving each model a role and a perspective:

• Diversity — planners reason from different angles (architect, implementer, researcher, critic, creative).
• Scaffolding — planners produce a concrete route; your model executes it with full tools; reviewers give actionable feedback.
• Synthesis over voting — a judge resolves contradictions instead of averaging, then a final pass integrates the consensus.
• Cost — a panel of open models can rival a frontier model for a fraction of the price per turn.

The pipeline

Fusion runs four stages around a single user request. The planning, review, judge and synthesis stages run server-side on the Omni Loop Portal — only the working stage runs on your machine, where your model uses your real tools.

1. Planning — 1–5 planner models analyse the request in parallel on the portal and output a structured plan: an ordered route of tool steps, concrete code blocks, risks and open questions.
2. Working — your current chat model receives those plans and does the real work with the full toolset (files, shell, browser, image generation, and more). This stage runs on your machine and is fully visible — you watch the tool calls stream.
3. Review — 1–5 reviewer models critique the draft from distinct angles (correctness, completeness, clarity, safety, delivery). Each returns a verdict — APPROVE, REQUEST_CHANGES or REJECT — with a score and exact fixes. If the judge asks for tool-bearing changes, one revision pass runs on your machine.
4. Synthesis — the agent applies every mandatory change, resolves contradictions, integrates unique insights and delivers one clean final answer. It never mentions the panel — you just get the result.

Turn it on

In any Clioloop surface, run /fusion for an interactive picker, or configure it directly:

Modes: auto gates out trivial requests, fast uses just the first advisor and reviewer, and full runs the whole panel.

Session hygiene

Treat each Fusion run as a clean-room collaboration. Begin from a fresh conversation so planners see only the current task, and reset after the final answer before starting another Fusion job. Clioloop shows a persistent reminder after successful Fusion runs, but it does not auto-reset your session because you may still want to read or save the final response.

Persisting a config

Fusion settings live in ~/.clio/config.local.yaml so they apply to every session:

Safe by construction

Planners and reviewers are read-only at the schema level — the write, shell and browser tools are not even in their toolset, so they can research and critique but can never touch your files or run commands. Only your main model acts, and you can watch it. See Security & tokens for more.