Ultron’s execution engine is a custom-built sandboxed runtime that runs in isolated cloud environments. It’s what makes long, complex, multi-step tasks possible — not a 3-minute timeout, but a 50-minute window with a full toolchain available from the start.
What the sandbox provides
Sandbox session starts
├── 50+ built-in tools loaded
├── 70+ slash commands available
├── 50+ MCP servers pre-configured
│ ├── Fixed: Brave Search, Puppeteer, GitHub, Apify, Tavily, Fetch, Filesystem, Memory
│ └── User-connected: Notion, Slack, Stripe, Linear, HubSpot, Figma, and more
├── Full filesystem access (read, write, execute)
├── Full browser automation (Playwright via Puppeteer MCP)
└── 50-minute runtime window
Tool execution pipeline
Every tool call follows the same pipeline:
Tool call received from Kimi K2
↓
validateInput() — Zod schema validation against tool definition
↓
checkPermissions() — Rule matching against allowed tool set
↓
call() — Execute the tool (API call, shell command, browser action, etc.)
↓
onProgress() — Stream live update to chat UI via SSE
↓
result() — Formatted output returned for next loop iteration
Parallel tool execution
When Kimi K2 returns multiple tool calls in a single turn, they execute concurrently. Independent operations don’t wait for each other:
// Three searches fire simultaneously
const [companyData, ceoBackground, recentNews] = await Promise.all([
web_search("Acme Corp overview"),
web_search("Acme Corp CEO background"),
web_search("Acme Corp news 2026")
]);
Budget control
Large tool outputs are handled gracefully:
- Results exceeding 30,000 characters are persisted to disk; the file path is returned instead
- The engine tracks token accumulation across all turns
- Compression triggers automatically before context limits are hit
- Sandbox sessions with heavy output (browser screenshots, large scraped pages) manage their own storage
Sandbox vs API mode
| API Mode | Sandbox Mode |
|---|
| Runtime limit | 3 minutes | 50 minutes |
| MCP servers | Selected subset | 50+ pre-loaded |
| Browser access | Via Browserbase API | Full Playwright |
| Filesystem | None | Full read/write |
| Plans | All plans | Growth, Scale |
Complex multi-step tasks — building a full competitive report, running a lead enrichment batch, generating a content calendar with research — run in sandbox mode to take advantage of the extended runtime.
