Kimi Agent Swarm tested in the cloud: 16-hour autonomous training run with self-recovery
solyarisoftware · x · 2026-09-18
Alessio Pomaro, Head of AI at Search On Media Group, ran two fully cloud-based tests with Kimi's multi-agent mode Agent Swarm (Moonshot AI), where a coordinator decomposes work across parallel sub-agents.
- Test 1: building a training environment with Google's ToolGrad — no GPU, 3GB RAM, flaky network and repeated filesystem wipes. The swarm finished anyway: the fine-tuned small model went from 0% to 76% valid syntax, 58% function recognition, 36% valid executions over 16 hours, much of it spent recovering from resets.
- Test 2: evaluating Google TimesFM 2.5/3.0 on Walmart M5 data with LightGBM baselines and multi-period backtests. Key finding: fine-tuning looked helpful at first but its edge vanished under rigorous 8-window validation — TimesFM 3.0 zero-shot averaged 7.57% WAPE vs 7.59% fine-tuned, LightGBM 8.x.
The standout observation is the autonomy level: recovering from errors and switching strategies mid-task.
More from coding & agent
- Salesforce launches Trusted Enterprise AI Harness to unify agent context, governance and security — emmanuelvivier · 2026-09-18
- Running Codex, Claude and Pi Agents Safely: gVisor Sandboxes Plus tart macOS VMs — craigbalding · 2026-09-18
- EvalSeal: open-source tool shows LLM judges flip verdicts on 5 of 20 borderline eval cases — Fit_Fortune953 · 2026-09-18
- Armin Ronacher floats replacing MCP with codemode + OpenAPI + RAG over API docs — mitsuhiko · 2026-09-18
- Obsidian Starter Kit v4 ships with MCP server, osk-cli and ~375 specialized AI skills — dSebastien · 2026-09-18
- Retrying LLM Requests Isn't Always Safe: Gateway Policies for Partial Streams and Side Effects — Rama_Surasani_ · 2026-09-18