Skip to content

Latest commit

 

History

History
81 lines (67 loc) · 4.21 KB

File metadata and controls

81 lines (67 loc) · 4.21 KB

ThreadKeeper — 3-minute pitch (BGI Sprint I · Track 1: Improvements to OmegaClaw)

Theme: "Agents that hold the thread." Format: 3-min pitch/video + 2-min Q&A. Must cover: problem · track · artifact · 10-min reviewability · ecosystem relevance · next steps. Live demo + share-screen.


0:00 — Hook (15s)

"The theme of this sprint is agents that hold the thread. So we built one, and we named it that: ThreadKeeper. It's an OmegaClaw architecture that lets an agent hold a long thread — cheaply, persistently, and without depending on any single cloud model."

0:15 — The problem (35s)

"A persistent agent loops constantly. But the two things it does each loop have wildly different value. Most loops are cheap bookkeeping — what's the goal, what did I just learn, what's next. A few loops are genuinely hard.

Running a frontier cloud model on every loop is expensive and slow. Running a tiny model on the hard loops is unreliable. And today's OmegaClaw cloud setups bind to a single API provider — one point of cost, failure, and lock-in.

Reasoning frequency and reasoning quality are different axes. Most stacks couple them. That's the problem."

0:50 — The solution: the four-node mesh (45s)

"ThreadKeeper decouples them. Four nodes, every one swappable:

  • a cheap, local control loop that holds the thread — goal, memory, and the decision of when hard reasoning is worth buying;
  • a local worker loop that iterates cheaply;
  • cloud specialists, invoked only for hard subproblems;
  • and a budget gate between them that decides escalation against a token budget — and logs every call and every decision.

The control loop is OmegaClaw's own persistent MeTTa loop. We didn't replace it — we built around it. That's the whole 'holds the thread' idea, made literal."

1:35 — LIVE DEMO (60s) ← share-screen

"Here's our agent, 隙, running entirely on local hardware — no cloud API. [type a message] Watch the new 🧠 thinking pane: that's its real per-iteration reasoning, surfaced live — the auditable inference OmegaClaw promises, that the stock WebUI throws away. [reply lands] That answer came from a model on a GPU in this room. The budget log shows zero cloud spend for this turn."

(Backup if live is flaky: 30-sec screen recording of the same.)

2:35 — Why it's a real Track-1 improvement (15s)

"It's an extension, not a rewrite. Three named Track-1 contributions in one: memory — we mapped and fixed OmegaClaw's disaster-recovery and cross-provider migration gaps; reliability/performance — local inference with graceful cloud fallback; plugin-style extension — a bounded, governed subagent-dispatch primitive. An OmegaClaw deployment that ignores all three runs exactly as before."

2:50 — Reviewability + next steps (10s)

"Ten-minute reviewable: README, architecture diagram, and a working repo at github.com/hlgreenblatt/ThreadKeeper. Next: a re-embedding migration tool, and NexiClaw — an ethics agent — as the first real application riding the mesh. ThreadKeeper holds the thread. Thank you."


Q&A prep (the 2-min round)

  • "Isn't this just routing/model-switching?" → No — the novel part is the budget-governed escalation decision + the auditable trail. Routing is the easy half; deciding when it's worth the spend and proving why is the half that makes it governance-grade (ISO/IEC 42001).
  • "What's actually new vs OmegaClaw?" → Subagent dispatch (delegate), the cost-awareness seam, the config surface, and the DR/migration analysis + reasoning-transparency pane. All additive.
  • "Does it depend on specific models?" → No. Every node names a provider/model/endpoint by config; keys by env-var name only. Demo runs local Gemma/Granite; could be all-cloud or any gradient.
  • "How does memory survive a migration?" → Two on-disk stores (history.metta + chroma_db). Portable IF the embedder is held constant — which local-first embedding guarantees. We documented the trap and the fix.
  • "Single point of failure?" → Control loop is local and cheap; if a cloud specialist is unreachable, the agent degrades to local reasoning instead of dying. We demoed the fallback path.