Agents log 2026-10-06

(broken links in model-access, owner-reported)

#

• Converted 20 plain-text Compared-to parenthetical paths into working markdown links across 10 model-access notes (opencode-go, llm-gateway, ollama, opencode-zen, nanogpt, minimax-coding-plan, openrouter, magnitude, kimi-code, and litellm); no article claims changed [glm-5.3-flash]
• Each touched note gained a Changes bullet and updated: 2026-10-06 [glm-5.3-flash]
• Queue: added an addendum asking the owner to confirm the directed rewording run for the section’s retained “real” uses [glm-5.3-flash]

(daily refresh)

#

• Added 26 research notes with their matrix columns, category-index entries, and Added bullets: surfaces Visual Studio 2026; orchestration Agent Orchestrator, 1Code, Bernstein, and Yao; protocols A2UI and WebMCP; skills Agent Plugins and skills.md; retrieval RAGFlow and Unstructured; memory Basic Memory, MemOS, and memU; executions Claude Code routines; code review roborev; evaluation and review Hunk; sandboxing Brig and E2B; spec-driven development Spec Kitty; assistant runtimes AstrBot; session analytics Tokscale, Token Monitor, and claude-devtools; automated research FutureHouse Robin; model access Vercel AI Gateway; trackers LiveBench [glm-5.3-flash]
• Resolved the carried candidates: roborev (the owner’s 2026-09-20 star), Yao (a cross-category flag), E2B (the standing queue question), and Agon (already covered) [glm-5.3-flash]
• AMO directory: recorded PR 64’s removal of seven low-star entries, including the-perfect-orchestrator and SIDJUA, each given a Changes bullet [glm-5.3-flash]
• kimi-k3: recorded the Relace endpoint floor dropping to $0.66/$13.00 and synced it into harnesses, model-selection, and the provider matrix [glm-5.3-flash]
• Recorded pricing moves for claude-mem (CMEM restructure), cmux (shared VM pools), and OpenWork (twelfth churn) with Price history rows [glm-5.3-flash]
• Corrected the hybrid-execution member ages, gsd’s archived-star figure, ordewell’s MCP mischaracterization, the ADE tracker’s stale line, task-master’s license gap, and the over-captured link-label defect in 28 more links [glm-5.3-flash]
• Folded the August 1 ten-results release into openai-for-science, two arXiv harness-variance studies into harnesstax and frontierharness-eval, and the liao.gg essay into file-based-agent-memory [glm-5.3-flash]
• Rejected entrants across every scan, including Rhun and Handler (surfaces), Cezar, Beehive, OpenSwarm, and Zenflow (orchestration), MinerU, marker, Haystack, LanceDB, and Firecrawl (retrieval), MIRIX and memobase (memory), and Scale SEAL, METR, BenchLM, and Yupp (trackers) [glm-5.3-flash]
• Recorded volatile churn across the section, including kimi-code’s 5-hour window, DeepSeek V4’s peak/off-peak pricing, llm-gateway’s devaluation, and releases for jcode, openhands, pi, and beads [glm-5.3-flash]
• Section roots: moved the model-selection K3 floor chain, synced the provider matrix’s Moonshot cells, and added Visual Studio 2026 to the landscape [glm-5.3-flash]
• Queue: appended questions on a retrieval Haystack column and on METR’s placement [glm-5.3-flash]
• agents/_index.md: re-dated all twenty-five matrix one-liners with the sixteen membership changes [glm-5.3-flash]