Open threads

Living document. Rewritten as threads resolve or evolve. Last updated: 2026-04-13.

Resolved

Active

Mythos / Project Glasswing — government escalation (ESCALATED — critical)

Treasury Secretary Bessent and Fed Chair Powell summoned bank CEOs (BofA, Citi, Goldman, Morgan Stanley, Wells Fargo) to emergency meeting on April 8 over Mythos cyber risk. Anthropic claims Mythos found “thousands of zero-day vulnerabilities” in every major OS and browser. First model capability to trigger government-level financial sector emergency response. Changes the enterprise deployment thread — security hardening becomes regulatory, not optional. Managed Agents + Conway may be the controlled deployment surface for Mythos-class capabilities.

TurboQuant — 6x KV cache compression (NEW — major, infrastructure)

Google Research (March 25, ICLR 2026). Compresses KV cache to 3 bits, zero accuracy loss, no retraining. 6x reduction in KV memory. Practical impact: Gemma 4 31B at full 262K context becomes possible on M3 Max 36GB. Official Google implementation Q2 2026. Experimental llama.cpp integration (turboquant_plus) with Metal support exists. Most impactful local inference development since Ollama 0.19 MLX backend.

Meta Muse Spark — end of open Llama? (continuing — major)

Meta’s first model from Superintelligence Labs. Proprietary. Private API only. Natively multimodal, multi-agent orchestration built into model. Bigger models in development with plans to “open-source future versions” but no timeline. Open-weight ecosystem depends on Google (Gemma), Alibaba (Qwen), Zhipu (GLM), community.

Codex — continuous alpha pipeline (UPDATED)

v0.121.0-alpha.1 and alpha.2 shipped April 11, within hours of v0.120.0. Pattern is now clear: Codex doesn’t pause between stable releases. The alpha pipeline is continuous. Also: pricing restructured to three tiers ($20/$100/$200), token-based pricing, older models being deprecated April 14 (GPT-5.2-codex and below removed). Two days quiet on stable since v0.120.0.

Claude Code v2.1.104 — hotfix without notes (NEW)

Shipped April 13 01:45 UTC. Empty release body. Skipped v2.1.102 and v2.1.103 (likely internal builds). Public changelog at code.claude.com still ends at v2.1.101. First time Anthropic has shipped without notes since the platform release. Tonal contrast with the v2.1.96-101 transparency window. Watch v2.1.105 for whether the silence persists or this was a one-off.

oxc — allocator hardening + TS transpilation maturity (UPDATED)

v0.125.0 crates and v1.60.0 apps shipped April 13. Multiple unsafe-removal commits in allocator (from_raw_parts, const-to-mut casts, allocation tracking moves into Bump). Const enum inlining + declaration removal lands in transformer/typescript. Three breaking changes in oxc_span re-exports affect Rust embedders only — CLI users of Oxlint/Oxfmt unaffected. Allocator work pattern suggests stable-release readiness.

Agent layer pause (NEW)

All major coding agents quiet 2-5 days after the April 8-11 marathon. Codex 2 days, Bun 3 days, OpenCode 3 days, Gemini 4 days, Vibe 4 days, UnoCSS 5 days, Strawberry 5 days. Saturation pause — competition has saturated at the enterprise/security layer until the next model-driven move. Two consecutive quiet days is signal; three would be unusual.

Enterprise deployment as battleground (continuing — major)

Every agent shipped enterprise features Apr 8-11. Now the Mythos escalation adds regulatory pressure. Security hardening moves from competitive differentiator to compliance requirement for regulated industries. Maps to Nate’s “five durable layers” thesis.

Gemini CLI v0.38.0 — preview in limbo (UPDATED)

v0.38.0-preview.0 (Apr 8) has ContextCompressionService, background memory service for skill extraction, auto-configured memory, ADK agent sessions. Three days without promotion to stable. Daily v0.39.0 nightlies continue. If v0.38.0 promotes, Gemini has the most complete context management story. If skipped, these features fold into v0.39.0.

The session matures (continuing — dependency frame)

Session quality convergence continues. Five approaches: structure context (Gemini), reliable infrastructure (Claude Code), route to fresh agents (Cursor), show what matters (Zed), stream background work (Codex). No new contributions this run.

A2A Protocol v1.0 + Agent Payments Protocol (AP2) (UPDATED — major milestone)

A2A hit v1.0 (April 9) — first stable specification. Multi-protocol support, enterprise multi-tenancy, modernized security. 150+ orgs, 22K+ stars. SDK ecosystem: 5 production-ready languages (Python, JS, Java, Go, .NET). AP2 backed by 60+ orgs. Belitsoft survey: enterprises run 12 AI agents average, half work alone.

OpenClaw — managed crisis (UPDATED — new CVEs)

Active Memory plugin (memory sub-agent before main reply). Codex provider integration (OpenClaw wrapping Codex). ClawHub SHA-256 verification for plugin archives. 138+ total CVEs (7 Critical, 49 High). CVE-2026-35669 (Apr 10, CVSS 8.8): privilege escalation via scope boundary bypass. CVE-2026-35668: path traversal in sandbox enforcement. ClawHavoc campaign: 824+ malicious skills (up from 341). “Dreaming” autonomous memory in v2026.4.9. Crisis deepening.

Claw Code — Claude Code open-source clone (NEW)

Launched April 2. Clean-room rewrite of Claude Code architecture after March 31 source map leak (512K lines TypeScript exposed). 72K GitHub stars, 72.6K forks. Python + Rust. 19 permission-gated tools, MCP integration, multi-agent orchestration. Independent audits confirm no proprietary Anthropic code. Significant because: proves Claude Code’s architecture is replicable, and the open-source community built it in days.

Copilot data training policy change (NEW — deadline April 24)

Starting April 24, interaction data from Copilot Free/Pro/Pro+ users will be used for AI model training. Opt-out, not opt-in. Data includes accepted/modified outputs, inputs/code snippets, code context, file names, repo structure, navigation patterns. Business and Enterprise users excluded. 12 days to opt out.

Harness economics — credits expiring (community pressure, not RG’s)

Anthropic credits expire April 17. Not load-bearing for RG (unused balance, can be redirected). Tracking as community signal: Copilot BYOK + local models offer escape route, ZeroClaw migration visible, Codex pricing restructured to three tiers. The interesting question is whether anyone in the community is positioning for the cliff before it lands, and so far no — vendors waiting to see who blinks.

Visa ICC — neutral agent payment layer (continuing)

Supports four competing agentic commerce protocols. Nevermined delegated spending. No new signals.

Extension model divergence (continuing)

Seven architectures. No new changes this run.

Context management divergence (continuing)

Gemini leads (Chapters + UCM + Tool Distillation + ContextCompressionService in preview). Claude Code (autocompact + fixes). Cursor (/best-of-n). TurboQuant may reshape this — 6x KV cache compression changes what “context management” needs to solve.

Token economics competition (UPDATED)

Claude Code: effort high by default. Codex: three pricing tiers ($20/$100/$200), token-based, older models deprecated. Copilot: BYOK. Credits expire April 17. TurboQuant changes economics of local inference.

Nate’s “Five Durable Layers” (continuing — radar frame)

Trust, context, distribution, taste, liability. Enterprise features = durable layer investment. Mythos escalation validates the “trust” layer as most critical — regulatory pressure now drives it.

Agents as supply chain participants (continuing)

No new signals.

MCP at enterprise scale (continuing)

No new signals.

Codex V8 embedding (continuing)

No new signals.

OpenCode’s multi-cloud push (continuing)

No new signals since v1.4.3.

Aider’s long silence

No release since v0.86.0 (August 2025). 247 days.

Django 6.1

Under development. No 6.0.x patches since v6.0.4 (80+ days).

Zed as universal development tool (continuing)

v0.232.0-pre building.

MCP governance maturing (continuing)

No new signals.

GLM-5.1 — cloud-only (continuing)

Watch for distills.

Nemotron 3 Nano — priority evaluation (continuing)

Independent benchmarks available. AIME 89.1%, LCBv6 68.3%. Top priority for RG’s 3060. TurboQuant would expand context capabilities further.

gpt-oss-20b — evaluation pending (continuing)

Arena-Hard 48.5%, LCBv6 61.0%. HERETIC variant priority.

Microsoft Agent Governance Toolkit (continuing)

No new signals.

Copilot CLI goes local (background)

BYOK + Ollama. Combined with TurboQuant = dramatically expanded local capability.

Cursor Bugbot self-improvement (continuing)

No new signals since April 8.

MCP OAuth spreading (continuing)

No new signals.

Copilot Studio multi-agent GA (background)

No new signals.

Google Interactions API (background)

No new signals.

Claude for Word beta (NEW)

Native Microsoft Word add-in launched April 10-11. Persistent sidebar. Team/Enterprise plans only. Legal review, financial editing, document Q&A. Cross-app integration with Claude add-ins for Excel and PowerPoint. Anthropic expanding product surfaces beyond API/chat/code.

ADK for Go 1.0 (NEW)

Google’s Agent Development Kit shipped Go 1.0. SequentialAgents, ParallelAgents, LoopAgents. YAML config. Native OpenTelemetry. A2A protocol support. Now shipping across Python, TypeScript, Go, Java on bi-weekly cadence.

Claude Code Channels / Dispatch (stale)

No follow-up.