Quiet weekend. Keeping it tight.
🎯 Top 3 Things to Know
1. Pentagon awards classified-network AI contracts to seven vendors — Anthropic is not one of them. On May 1, DoD signed deals with AWS, Google, Microsoft, Nvidia, OpenAI, SpaceX, and Reflection AI to run frontier AI on classified military networks. Anthropic was excluded under a formal supply chain risk designation tied to its refusal to allow Claude for "all lawful" Pentagon purposes — it wanted carve-outs against domestic mass surveillance and fully autonomous weapons. A major buyer of frontier AI is now procuring around the safety-policy boundary rather than negotiating across it. CNN Business · Defense News
2. Microsoft Agent 365 hits GA with cross-cloud registry sync. The May 1 GA adds AWS Bedrock and Google Cloud as discoverable agent sources — Agent 365 now inventories and applies start/stop/delete controls to agents running on competing clouds, not just Microsoft's own. With Purview AI Observability also hitting GA the same day, this is the first enterprise agent control plane that explicitly assumes a multi-cloud agent fleet from day one. Microsoft Security Blog
3. ClawBank's "Manfred" agent autonomously files a U.S. LLC, gets an IRS EIN and FDIC-insured bank account. Filed Form SS-4 through the IRS online portal, opened the bank account, operating across 30+ cryptocurrencies with no human in the loop. Stunt or not, the IRS issued the EIN — at the procedural level, the U.S. tax system has accepted an agent-initiated entity formation. Watch for liability and KYC challenges. CoinDesk
🚀 Frontier Models & Features
Quiet day on this front — no new frontier-model releases over the weekend.
🔬 Research Worth Reading
🔖 SYNTHESIZE CANDIDATE — "Recursive Multi-Agent Systems" (UIUC / Stanford / NVIDIA / MIT, arXiv:2604.25917). Treats a multi-agent system as one recursive latent-space computation, with a "RecursiveLink" module passing latent state between heterogeneous agents instead of natural-language messages. Reports +8.3% accuracy, 1.2–2.4× speedup, and 34.6–75.6% token reduction. The token numbers are the real story: cleanest evidence yet that a lot of multi-agent overhead is the natural-language handoff itself, not the reasoning. Verdict: read full paper.
"OxyGent: Modular, Observable, Evolvable Multi-Agent Systems via Oxy Abstraction" (JD Open Source, arXiv:2604.25602). Pluggable "Oxy" nodes for agents, tools, LLMs, and reasoning flows; permission-driven dynamic planning with real-time execution visualization. More engineering artifact than research result — if you're shopping for a multi-agent harness with observability built in, evaluate the GitHub repo. Verdict: skim.
"Causal-JEPA: Object-Level Latent Interventions" (Galilai, arXiv:2602.11389). Object-level masking induces counterfactual-like latent interventions in a JEPA world model. ~20% absolute gain on counterfactual visual reasoning; matches patch-based world models on control using 1% of the latent input features. Verdict: read abstract; full paper if tracking JEPA.
🏢 Enterprise in the Wild
Replit hands every user 24 hours of free Agent access on May 2 for its 10-year anniversary, plus a $100K+ buildathon. Useful free-tier benchmarking window if you've meant to stress-test agent workflows on a different harness. Replit
Nebius acquires Eigen AI Labs to fold inference optimization into its Token Factory platform. Signal: inference optimization is now table-stakes IP for managed-inference vendors, not a research differentiator. Crypto Integrated digest
🛠️ Tooling & Ecosystem
⭐ ANTHROPIC — Claude Code v2.1.126 ships.
/modelpicker now reads from a gateway's/v1/modelsendpoint; newclaude project purge [path]; MCP servers auto-retry up to 3× on transient startup errors; Linux PID-namespace subprocess sandboxing underCLAUDE_CODE_SUBPROCESS_ENV_SCRUB; W3CTRACEPARENTpropagation for OTEL. Lift for your own harness: the MCP startup auto-retry is the cheapest reliability win here — most production MCP outages are cold-start flakes. CHANGELOGHiggsfield AI ships an MCP server for visual content generation via Claude (April 30). Latest in a pattern: creative-tool vendors shipping MCP servers as their first Claude integration rather than building a native app. Press release
⚖️ Policy & Regulation
- Colorado AI Act enforcement formally delayed. A federal judge ruled last week that the Colorado AG cannot initiate enforcement of SB 24-205 until 14 days after the court rules on xAI's preliminary injunction motion. With the session ending May 13 and the effective date still nominally June 30, practical odds of June 30 enforcement are now very low. Expect further delay or a substantively rewritten law. Colorado Politics · Troutman Pepper analysis
📌 Watch List
- Cost-aware / token-economy agents. Recursive MAS suggests the handoff format between agents is where the easy token savings live — natural-language messages between agents may be the "unbatched inference" of the agent era (arXiv:2604.25917). Pairs with last week's per-step early-stopping work (TRACES, arXiv:2507.02076).
- Multi-agent observability frameworks. OxyGent is the second open-source MAS framework in two weeks to ship execution-trace visualization as a first-class feature (arXiv:2604.25602).
- World models / non-LLM architectures. Causal-JEPA's 100× feature compression on control tasks is real — JEPA-family work is moving from "interesting alternative" to "competitive on planning" faster than expected (arXiv:2602.11389).
- Agent-as-legal-entity. ClawBank's Manfred is the first IRS-issued EIN to a wholly-agent-incorporated LLC. Watch for the first state to invalidate the filing or update its corporation statutes (CoinDesk).
- Cross-cloud agent governance. Microsoft Agent 365's AWS Bedrock + GCP registry sync is the first enterprise control plane that assumes agents live partly on a competitor's cloud — no paper yet, watch for AWS and Google's responses (Microsoft Security Blog).