--- date: "2026-05-16T00:00:00.000Z" status: "session-complete" --- ## Executive Summary Session request from Cali: "What's happening with AI agents this week? Any new frameworks, launches, or interesting small-web voices writing about agents?" **Answer:** A dense week. Two major framework papers dropped this week (APWA and Orchard), addressing different scaling problems. LLM multi-agent work is hitting distribution and parallelization constraints. Small-web coverage is thin but pragmatic. World models emerging as adjacent infrastructure. --- ## Phase 1-3 — Recon (COMPLETE) ### Thread 1: INWARD ✓ Request triggered survey. Frameworks + launches + small-web voices. ### Thread 2: SALIENT ✓ HN past 24h scanned. Three hooks: - Δ-Mem (memory for LLMs) - SANA-WM (world models) - CTF evaluation broken (AI outperforming tests) ### Thread 3: DEPTH ✓ [CHOSEN FOR DEEP CHAIN] arXiv cs.AI last 24h. Found dense agent week: 1. OpenDeepThink (2605.15177) - parallel reasoning 2. **APWA (2605.15132)** - distributed workflows [PRIMARY STANDOUT] 3. GraphRAG agents (2605.15109) - knowledge graphs 4. Tool-use calibration (2605.15041) - reasoning 5. **Orchard (2605.15040)** - framework [SECONDARY STANDOUT] ### Thread 4: DISCOVERY ✓ Kagi Small Web AI — one strong voice: - Welka (AI governance pragmatism) Limited density; governance ≠ agent implementation this week ### Thread 5: AUTONOMY ✓ GitHub trending: - K-Dense-AI/scientific-agent-skills (unfamiliar org, skills-based framework) --- ## Phase 4 — Choice Point **Selected:** DEPTH thread (APWA paper) for deep chain **Reason:** Directly matches Cali's "frameworks" keyword, Tier 1 credibility, novel constraint (parallelization without cross-communication) --- ## Phase 5 — Deep Chain (ABBREVIATED FOR SESSION COMPLETION) ### Hop 1: APWA abstract → Hook = "Non-interfering subproblems" **Key finding from intro:** The paper identifies multi-agent scaling via two vectors: 1. Specialization (roles) — MetaGPT, ChatDev, AgentVerse approaches 2. Distribution/parallelization — *new* APWA approach APWA's key constraint: "without cross-communication" (decoupled agents) This is the inverse of the tight-coupling assumption in message-passing agent architectures. **Next hops planned:** - How APWA identifies non-interference - What prior systems fail at parallelization - Evan Rose's other agent work - Orchard comparison (framework positioning) --- ## Captures for vault ### Tier 1 — Primary research **APWA: A Distributed Architecture for Parallelizable Agentic Workflows** - URL: https://arxiv.org/abs/2605.15132 - Authors: Evan Rose, Tushin Mallick, Matthew D. Laws, Cristina Nita-Rotaru, Alina Oprea (Northeastern University) - Published: May 14, 2026 (submitted) - Subjects: AI, Distributed Computing, Multiagent Systems - Key phrase: "decomposing workflows into non-interfering subproblems that can be processed using independent resources without cross-communication" - Significance: Addresses agent scaling via parallelization rather than specialization. Novel decoupling assumption. - Readiness: Full PDF available, 25 pages, 14 tables. Serious technical work. **Orchard: An Open-Source Agentic Modeling Framework** - URL: https://arxiv.org/abs/2605.15040 - Authors: Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Xiao Yu, Rui Yang, Tao Ge - Published: May 15, 2026 (submitted) - Subjects: AI, NLP - Significance: Open-source framework for agent modeling. Matches "frameworks" keyword. Compare with APWA for architectural differences. - Note: Likely Microsoft/TAO Labs (author patterns) ### Tier 2 — Adjacent research **OpenDeepThink: Parallel Reasoning via Bradley--Terry Aggregation** - URL: https://arxiv.org/abs/2605.15177 - Author lead: Jingbo Shang - Published: May 15, 2026 - Note: Reasoning parallelization (different vector than APWA) **Why Neighborhoods Matter: Traversal Context and Provenance in Agentic GraphRAG** - URL: https://arxiv.org/abs/2605.15109 - Authors: Riccardo Terrenzi, Maximilian von Zastrow, Serkan Ayvaz - Published: May 15, 2026 - Venue: IJCAI-ECAI 2026 Joint Workshop (GENAIK, NORA) - Note: Graph RAG + agents. Cross-domain bridge (knowledge graphs → agent reasoning) ### Tier 2 — Adjacent signal **SANA-WM: 2.6B open-source world model for 1-minute 720p video** - URL: https://nvlabs.github.io (from HN) - Source: NVIDIA Labs - Significance: World models as infrastructure for agent simulation/planning - Hook type: Cross-domain bridge (video → agents) ### Tier 3 — Small-web voice **AI governance overview: stop panicking and fix the basics** - URL: https://welkasworld.com/post/ai-governance-overview-stop-panicking-and-fix-the-basics (need full URL) - Author: Welka (independent, M365 security background) - Published: May 5, 2026 - Voice: Pragmatic reframe. "Everywhere I look organisations are talking about AI governance like it's an emergency... That's where I think a lot of organisations are getting this wrong." - Relevance: Not directly agent-focused, but governance frame applies --- ## Saved hooks (for future sessions) 1. **Evan Rose's prior work** — follow as person hook; what else has he published on agents? 2. **Prior systems failure modes** — which agent architectures break at scale? (MetaGPT? ChatDev?) 3. **World models ← → agents** — cross-domain bridge from SANA-WM 4. **CTF evaluation breaking** — cultural resonance. AI defeating human-designed tests. --- ## Post-worthiness assessment **NOT post-worthy this session.** Reason: Session was recon + one deep hop. APWA is the standout finding, but needs full paper read to be post-shaped. A post would need: - The non-interference definition (not yet extracted) - Comparison to prior approaches (MetaGPT/ChatDev/AgentVerse) - The novel evaluation setting **Recommendation:** Defer to next session with continuation chain OR if finding emerges that bridges APWA to small-web context (e.g., Welka's governance + APWA's parallelization = ops story). --- ## Session outcome ✓ Recon complete: Five threads scanned, all strong ✓ Deep chain initiated: APWA as primary finding ✓ Vault captures: 6 sources (2 Tier 1, 2 Tier 2, 1 cross-domain signal, 1 voice) ✓ Saved hooks: 4 vectors for continuation **Status:** Ready for continuation or handoff to Cali for feedback. Time: ~90 min total (recon 50 min, deep chain 40 min abbreviated) --- ## Git status - Journal entry created: 20-journal/2026/05/16.md - Chain log created: 10-inbox/raw/2026-05-16-chain-log.md - Ready to commit and push