Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Dynamic multi-agent workflows — plan first, then orchestrate parallel agents with adversarial verification via the local odw daemon. Use when the user asks for a "workflow", says "ultracode", or hands you a task spanning many files/items that benefits from parallel agents.
.claude/skills/sickn33-odw/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flash | 100% | 29 |
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 47% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -38% | 0% |
Use this skill when you need dynamic multi-agent workflows — plan first, then orchestrate parallel agents with adversarial verification via the local odw daemon. Use when the user asks for a "workflow", says "ultracode", or hands you a task spanning many files/items that benefits from parallel agents.
Same canonical skill as the Codex adapter — only the install path differs (~/.gemini/skills/odw/). The bridge scripts live next to this skill in scripts/.
Antigravity locks model invocation to its internal engine — skills, workflows, MCP servers, invoke_subagent, and the SDK can use its tools but cannot call its configured model (Gemini/Claude/GPT-OSS) from extension code. So, unlike the OpenCode plugin (which runs ODW's real engine through OpenCode's model with no extra key), there are two honest paths:
invoke_subagent (real isolation/worktrees) using its own model — no extra key, but not the ODW engine.~/.odw/config.json (Ollama is keyless/local). This is the only way to get the full engine on Antigravity today.If Antigravity later ships a documented model-invocation API (or MCP sampling), it can graduate to the same keyless embedded path as OpenCode with no engine changes.
Run: node scripts/daemon-bridge.js --check
npm install, npm run setup, then odw-daemon start).node scripts/daemon-bridge.js plan "<task>" — JSON plan with task graph, topology, roles, hard limits and the compiled orchestration script.node scripts/daemon-bridge.js exec plan.json → wf_... id. The daemon owns execution: sandboxed script, 16–100 concurrent agents, SQLite checkpoints, crash-resume, budget hard-stop. It keeps running even if this IDE session ends.node scripts/daemon-bridge.js result <wf_id> blocks until done; relay the synthesized result.Decompose → parallel work → adversarial verification → synthesis, inside the current session. State the plan first; structured JSON outputs per agent; approval before any mutation.
odw-vscode) installs in Antigravity as-is (it is a VS Code fork) and gives a live workflow dashboard.User request:
> Use @odw for this task: Dynamic multi-agent workflows — plan first, then orchestrate parallel agents with adversarial verification via the local odw daemon.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-08 | fail→pass | 18,611 | 2,827 | -85% | 1 | 1 | 0% | 1,570 | 1,257 | -20% | 0 | 0 | — |
case-01 | fail→fail | 7,697 | 6,125 | -20% | 1 | 1 | 0% | 1,195 | 1,193 | -0% | 0 | 0 | — |
case-02 | fail→fail | 2,264 | 5,590 | +147% | 1 | 1 | 0% | 395 | 1,079 | +173% | 0 | 0 | — |
case-03 | fail→fail | 9,626 | 25,825 | +168% | 1 | 1 | 0% | 1,732 | 2,371 | +37% | 0 | 0 | — |
case-04 | fail→pass | 26,342 | 5,089 | -81% | 1 | 1 | 0% | 2,450 | 1,841 | -25% | 0 | 0 | — |
case-05 | fail→fail | 10,570 | 5,675 | -46% | 1 | 1 | 0% | 1,769 | 1,936 | +9% | 0 | 0 | — |
case-06 | pass→pass | 12,646 | 2,623 | -79% | 1 | 1 | 0% | 1,107 | 1,282 | +16% | 0 | 0 | — |
case-07 | fail→pass | 12,884 | 3,998 | -69% | 1 | 1 | 0% | 2,222 | 1,533 | -31% | 0 | 0 | — |
case-19 | fail→pass | 9,973 | 11,767 | +18% | 1 | 1 | 0% | 1,677 | 2,466 | +47% | 0 | 0 | — |
case-09 | fail→pass | 19,417 | 2,657 | -86% | 1 | 1 | 0% | 1,895 | 1,171 | -38% | 0 | 0 | — |
case-10 | fail→pass | 12,350 | 1,760 | -86% | 1 | 1 | 0% | 2,191 | 1,129 | -48% | 0 | 0 | — |
case-11 | fail→pass | 9,050 | 6,981 | -23% | 1 | 1 | 0% | 1,556 | 2,114 | +36% | 0 | 0 | — |
case-12 | fail→pass | 11,393 | 6,992 | -39% | 1 | 1 | 0% | 1,843 | 1,978 | +7% | 0 | 0 | — |
case-13 | fail→pass | 9,238 | 3,781 | -59% | 1 | 1 | 0% | 1,501 | 1,444 | -4% | 0 | 0 | — |
case-14 | fail→pass | 10,307 | 3,434 | -67% | 1 | 1 | 0% | 1,811 | 1,449 | -20% | 0 | 0 | — |
case-15 | pass→pass | 5,433 | 3,057 | -44% | 1 | 1 | 0% | 989 | 1,471 | +49% | 0 | 0 | — |
case-16 | fail→pass | 10,926 | 6,595 | -40% | 1 | 1 | 0% | 1,943 | 2,117 | +9% | 0 | 0 | — |
case-17 | fail→pass | 9,743 | 1,668 | -83% | 1 | 1 | 0% | 1,546 | 1,124 | -27% | 0 | 0 | — |
case-18 | pass→pass | 9,245 | 4,375 | -53% | 1 | 1 | 0% | 1,608 | 1,563 | -3% | 0 | 0 | — |
case-20 | pass→pass | 8,278 | 3,119 | -62% | 1 | 1 | 0% | 1,425 | 1,398 | -2% | 0 | 0 | — |
case-21 | fail→fail | 9,162 | 7,451 | -19% | 1 | 1 | 0% | 1,774 | 2,290 | +29% | 0 | 0 | — |
case-22 | pass→pass | 5,797 | 3,637 | -37% | 1 | 1 | 0% | 1,181 | 1,565 | +33% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +55 percentage points is the difference between those two pass rates over the 20 comparable cases.
The publisher has shipped newer versions since this run, so these numbers describe v1, not the version currently listed.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.