A single shared model plays executor, workflow optimizer, and evaluator, searches for better multi-agent workflows, and fine-tunes on executions of the team it judged best. After two cycles with Qwen3.6-27B, score per output token reached 1.2 to 1.6x the base model on four research benchmarks, and the authors found the optimizer, not the evaluator, was the bottleneck; paper and code are open.
The environment, not the model, logs every file a terminal agent reads, tracks listed but unopened files, and blocks the first finish while requirements are still open; the authors found about one in five values an agent had already seen never made it into the deliverable. With the same Pi based harness it beat plain direct workspace agents and model kept TODO lists on Workspace Bench, TheAgentCompany, and OfficeQA Pro, lifting Gemini 3.8 Flash's Workspace Bench pass rate from 31.4% to 46.9%…
Both sessions join a named "bag" in the same repo, and a review request or reply lands in the other agent as a new turn through its own input, with no daemon, polling, or hooks and one shared local log. The MIT server needs Python 3.10+ and a Codex CLI with codex queue , installs with pipx install 'postbag[mcp]' , and is tested on macOS…
Instead of calling MCP tools one by one and dumping every full result into context, the model writes a single JavaScript program that calls the tools, filters and joins the output in an isolated Node process, and returns only the final answer. It needs Node.js 22.13+, keeps normal MCP permission checks, and installs with /plugin marketplace add gabe4coding/claude code mode…
By his napkin math, about 6B people have barely touched LLMs, 1 to 2B use free tier chatbots, roughly 20M see agents finish weeks long math and code projects from a prompt, and about 5,000 inside labs see swarms of agents working for weeks. He blames scale of ambition, capability peaking in verifiable domains like code, and access tiers, and warns that human review is starting to fall behind…
A daily OpenCode and Claude Code user says agents still run parallel subtasks one at a time, never hand gruntwork to cheaper models, write hard to read plans, stall overnight on trivial questions, and lack real OS level sandboxes. He lists table stakes he wants out of the box, including per task model routing, deterministic filesystem and network boundaries, and an AFK mode that decides after 30 minutes without a reply…
Spotted something we missed? Start a thread.