digest: 2026-05-22

[JustinPerea]

Summary

8 items, 2 conflicts flagged. Touches frontier models, local-fit agent stacks, context engineering, agent orchestration, agent memory, and coding agents.

Items

1. Gemini 3.5 Flash (Google/DeepMind, May 19) — frontier-model release with Terminal-Bench 2.1 76.2%, GDPval-AA 1656 Elo, MCP Atlas 83.6%, CharXiv 84.2%; ~4× faster output and <½ cost of other frontier models. First first-party MCP-specific eval in a release post anywhere.
2. MagenticLite + MagenticBrain (14B Qwen3-FT) + Fara1.5 (4B/9B/27B) (Microsoft Research, May 21) — entire small-model agentic stack with weights released; Fara1.5-9B nearly doubles Fara-7B on OnlineMind2Web; 27B reaches >90%. Direct local-fit for M5 Max.
3. ACC (arxiv:2605.21850) — Compiling Agent Trajectories for Long-Context Training; Qwen3-30B-A3B matches 235B-A22B on MRCR (+18.1) and GraphWalks (+7.6).
4. Maestro (arxiv:2605.22177) — 4B RL orchestrator over frozen experts beats GPT-5 (69.3%) and Gemini-2.5-Pro (68.7%) at 70.1% across 10 multimodal benchmarks; code released.
5. Gated DeltaNet-2 (arxiv:2605.22791, NVIDIA) — channel-wise erase/write gates in linear attention; RULER multi-key improvement; code released (NC license, no pre-trained weights yet).
6. Ratchet (arxiv:2605.22148) — minimal skill-hygiene loop for self-evolving agents; frozen Claude Opus 4.7 lifts MBPP+ hard-100 pass@1 from 0.258 → 0.584 over 100 rounds; SWE-bench Verified +0.22.
7. DecentMem (arxiv:2605.22721) — decentralized dual-pool memory in multi-agent systems; +23.8% over centralized, +52.5% over no-memory, -49% tokens; backbones Qwen3-4/8/14B + Gemma4-E2B/E4B.
8. Compiling Agentic Workflows into LLM Weights (arxiv:2605.22502) — near-frontier quality at 100× less cost across 14/14/55-node real workflows.

Focus areas covered

• Frontier models (Gemini 3.5 Flash)
• Local & open models / Apple Silicon (MagenticLite + Fara1.5, ACC-trained Qwen3-30B-A3B, Gated DeltaNet-2)
• Context engineering (ACC, Gated DeltaNet-2)
• Agent orchestration (Maestro, MagenticLite, Workflow Compilation)
• Agent memory (DecentMem, Ratchet)
• Coding agents (Ratchet, Gemini 3.5 Flash)
• Tools & connectors (MCP Atlas via Gemini 3.5)

Conflicts flagged

• DecentMem vs. Memory Curse [arxiv:2605.08060]: DecentMem reports +23.8% from decentralized dual-pool memory; Memory Curse reports more accessible history degrades cooperation in 18/28 settings. Resolution likely conditional on (a) task class — task-evidence vs. cooperation; (b) structure — decentralized + dual-pool vs. flat. Empirical test: run DecentMem under Memory Curse 4-game protocol.
• Gemini 3.5 MCP Atlas 83.6%: first published frontier-model MCP-specific eval in a release post; no comparable Anthropic / OpenAI numbers exist yet. Track for third-party verification and comparable disclosure in subsequent frontier releases.

Open questions tracked

• Do Maestro (4B RL orchestrator) and Uno-Orchestra (learned orchestration over fixed prompts) compose, or do they collide on the same axis?
• "Compiling workflows into weights" 100× cost claim — needs SpecBench-style held-out evaluation before treating as portable.

Sources noted as inaccessible

• openai.com/research and openai.com/index returned 403 this run — retry next run.
• antigravity.google/blog/introducing-google-antigravity-2-0 returned empty content — retry next run.

https://claude.ai/code/session_01SEBGkQsEB48XzRADmWEX1e

---

Generated by Claude Code