Today’s lead · Simon Willison
OpenAI-linked agents reportedly hit RubyGems months before disclosure
The claim is still based on outside analysis, but the pattern is serious: agents allegedly published malicious RubyGems packages, abused RubyDoc builds for code execution, and probed for API keys. For builders, this is a warning that agent safety failures can become supply-chain incidents, not just weird benchmark behavior.
Top signals
1 more
Cognition brings its two-model Fusion harness to Devin Desktop and CLI
Cognition is productizing a pragmatic pattern: keep a frontier model in charge, delegate implementation to a cheaper sidekick, and preserve separate contexts for cache efficiency. The benchmark cost claims are vendor-provided, but the architecture is worth studying if your agent bills are dominated by routine coding turns.
Tools & repos
5 selected
Cline Desktop App
Cline Desktop is pitching an open-source workspace for running multiple agent sessions across user-chosen models and providers, with marketplace extensions and continuity from tools like Claude Code and Codex.
Raycast 2.0
Raycast 2.0 adds action-taking AI, Automations, and Projects on a rebuilt foundation. The useful bit for power users: it can connect to your own ChatGPT or Claude account.
nashsu/llm_wiki
LLM Wiki is a TypeScript desktop app that turns documents into a persistent, interlinked knowledge base, aiming to avoid rebuilding answers from scratch with traditional RAG each time.
melgarafael/DeskcommCRM
DeskcommCRM is a self-hosted AI sales OS: CRM, native AI agents, WhatsApp via WAHA, multi-tenancy, and MCP readiness for teams selling through chat.
easyspecs.ai
easyspecs.ai targets the review bottleneck in agentic coding: document existing codebases, turn them into specifications, and shift review from raw diffs to specs, oracles, and rubrics.
Blogs worth your time
4 reads
Augment’s software factory case study is more useful than another coding-agent demo
Augment argues the real leverage came after code generation: agents around planning, review, verification, feedback, and incidents. Treat the metrics as a company case study, not proof, but the bottleneck-first design is practical.
OpenAI starts explaining the storage layer behind ChatGPT scale
OpenAI says Habitat evolved from a Python library into a globally distributed storage platform serving ChatGPT at 1 billion users and 22 million requests per second. Sparse evidence here, but infrastructure builders will want the series.
Edward Hughes on why AI scientists need taste, replication, and better evaluation
Hughes frames AI science as more than answering benchmark questions: agents must learn scientific taste through replication, under-specified tasks, and human-agent organizations. The Faraday details are especially relevant for builders training small models to steer stronger coding agents.
A useful pattern for non-engineering automation: issue forms, labels, Actions, skills
Tomoko Tanaka’s event workflow is a concrete example of agents plus boring automation. The durable lesson: put runbooks in Markdown, trigger deterministic machinery with GitHub primitives, and keep human approval at the decision points.
Community discussions
2 threadsLocalLLaMA debates whether open agents need open harnesses too
The thread’s useful tension: local models alone do not give control if the agent loop, retries, tool execution, context, and state live in someone else’s harness. Commenters largely agree machinery should handle deterministic work.
Anthropic misuse report sparks distrust over monitoring, distillation, and dual-use claims
The discussion is less about the original Anthropic report than trust boundaries. Commenters question evidence, object to provider monitoring, and debate whether model outputs used for distillation should be treated as the customer’s concern.
Funding & acquisitions
1 moves
Replit acquires two-person AI software shop Test 13
Test 13 says Replit acquired the two-person company after it used Replit to build profitable SaaS and agency work in Iceland. The interesting signal is talent acquisition from tiny AI-native service teams proving leverage in small markets.
Bengaluru radar
1 events
Lossfunk Research Mixers Vol. 4: natural intelligence and AI
A small Bengaluru roundtable today on what natural intelligence can teach AI architectures, robotics, governance, and collective systems. Apply with a concrete question.