Context Engineering with Multi-Agent Routing
The practical loop behind Project Transformers: choose the right context, route to the right agent or tool, preserve the outcome, and let the system get smarter without turning the prompt into a junk drawer.
The routing loop
Signal arrives
Telegram message, Fizzy card, cron wake, GitHub issue, PR check, browser page, or human nudge.
Select context
Load only the relevant memories, conventions, project files, and recent state needed for this decision.
Route intent
Pick the execution lane: Optimus, specialist agent, coding harness, first-class tool, cron, or human review.
Execute safely
Use scoped tools, repo rules, reversible changes, protected pages, and explicit boundaries for external writes.
Persist result
Commit code, update Fizzy, append memory, create PRs, add knowledge, or mark a blocker.
Feedback loop
Heartbeats, dispatcher state, status dashboards, and future retrieval turn the output into better future context.
What gets routed where?
| Signal | Context boundary | Route | Durable output |
|---|---|---|---|
| User ask | Current chat + relevant memory/docs only | Optimus or specialist agent | Answer, file change, PR, or Fizzy card |
| Fizzy card | Card description, steps, comments, repo conventions | Dispatcher → coding agent / Optimus / human | Branch, PR, card update, review note |
| GitHub issue | Issue labels, repo rules, CI, linked PRs | Wheeljack / Amp / Codex / OpenCode | Implementation PR with verification |
| Recurring check | Cron payload + current runtime status | Isolated OpenClaw job | Summary, alert, state file, or silence |
| Knowledge gap | Memory + QMD + source files | Search, retrieve, summarize, compound | Knowledge note or updated convention |
Fizzy demo path
Human creates/updates a task with steps and assignee.
Fuel, blockers, worktree, routing lane, runtime caps.
PR, comment, done step, or “needs judgment”.
For the Card #273 demo, start with System Architecture → Communication Architecture to show Fizzy cards, fizzy-pop events, direct messaging, and Telegram topics together, then jump to Ops Dashboard → Fizzy Dispatcher for the live query/fuel/route flow.
Speaker notes
Three punchlines
- Context engineering is routing. The system decides what context enters the turn before it decides what model answers.
- Agents are boundaries, not personalities. Each agent owns a domain, memory surface, and permission envelope.
- Memory is an output channel. If a result matters later, it must become a card, commit, note, or convention.
Where this sits in the existing Viz tour
Start
System Architecture for the full Autobot topology and communication map.
Deepen
Shared Conventions and Memory Management for context selection.
Prove
Operations Dashboard for live cron, fuel, dispatcher, and health posture.