IAF Super Chat
A special IAF build you talk to in plain language — it writes its own new agents through autonomous coding, and reviews and improves its own work.
Super Chat is an IAF shell: a thin, editable layer wrapped around the IAF engine that specializes the general-purpose framework into a conversational, self-extending platform. Three things define it: it talks with you, it codes new capabilities for itself, and it reviews and refines its own work.
One Microphone, Decentralized Hand-off
Most multi-agent systems route through a central orchestrator that calls sub-agents as hidden tools and stitches their answers back together. That star topology loses context at every hop and makes the "who is really talking to me" question murky. Super Chat uses a different model:
- One microphone. At any moment exactly one dispatch holds the mic and talks to you directly.
- Peer-to-peer hand-off. Any mic-holder can pass the mic to another dispatch, which then answers as itself, inheriting the complete shared conversation. The user sees one continuous turn: zero or more hand-offs, then exactly one answer.
- Single-track shared context. All dispatches in a session share one conversation track — the successor never re-reads from scratch and never asks you to repeat yourself.
It Codes New Capabilities for Itself
Describe a new agent in words, and a built-in planner + coder pair designs it and builds it autonomously — scaffolding a new dispatch inside a sandboxed write fence and syncing it live. A new dispatch is a new action capability, authored by the platform on request. Just ask the front-door agent to build something: it hands off to the planner, which refines a plan with you, then hands off to the coder, which builds and verifies it. This is the platform building itself.
It Reviews and Refines Its Own Work
A self_refiner dispatch inspects how a dispatch is running, diagnoses issues, optimizes prompts (auto-backing-up the old version), and produces debug plans — the platform improving itself.
The Dispatch Roster
| Dispatch | Holds mic? | What it does |
|---|---|---|
| chat | yes | The Architect — the default front door. Discusses concepts and architecture; entry point to knowledge-base exploration. |
| dispatch_building_planner | yes | Works with you to define what dispatch to build, producing a plan. |
| dispatch_coder | yes | Reads the plan and builds the dispatch inside a sandboxed fence, then syncs it live. |
| self_refiner | yes | Self-review: explores how a dispatch runs, diagnoses issues, or optimizes a prompt. |
| shadow_explore | tool | Read-only exploration of the knowledge base / codebase. |
| web_searcher | yes | Free-search web look-ups behind a fenced tool loop. |
Shell + Engine
You edit four zones — config_base/, prompt_base/,
tools_base/, and a human-authored knowledge base wiki/. A single
entry point syncs those sources into the engine (translating shell-root-relative paths into
engine-relative ones), then starts the server. Editing a prompt takes effect on the next turn;
changing a config requires a re-sync. Dispatches that write files or run commands do so inside
explicit fences — the only roots they may touch, with a per-token command allow-list.
Get It
GitHub: IntelligenismCommercialDevelopment-LLC/iaf_super_chat ↗
Derived from the AI Operability Edition of the Intelligenism Agent Framework — specialized into a conversational, self-extending, self-refining chat platform.