THE COLLECTION
EVERY SIGNAL, SORTED.
Search the whole shelf by topic, tool, term, or plain old curiosity.
United States, OpenAI, Coding Agents, AI Research, Scientific Productivity, Research Evaluation, Compute, Human Oversight, Safety and Recursive Self-Improvement
Coding agents finish nearly half of short research tasks. Measure the whole lab
OpenAI says its agents' estimated zero-intervention success on two-to-four-hour research tasks reached 46 percent in October. The result is striking, uneven and easy to overread. A completed coding task is one station in a laboratory, not the laboratory's output.
France, Europe, Mistral AI, Vibe Desktop, Coding Agents, Local Development, Cloud Sandboxes, Permissions, Credentials, Worktrees, GitHub, Background Processes, Prompt Injection, Audit Trails and Software Safety
Mistral put local and cloud coding in one desktop. Mark the boundary before the agent crosses it
Vibe Desktop puts local repositories and managed cloud coding behind one interface. The continuity is convenient. The authority is not continuous. A short environment receipt should name the checkout, credentials, network, tools, processes, evidence and final destination before work moves between a developer machine and a cloud sandbox.
Open Source AI, Coding Agents, Developer Tools, Local Computers, Tool Permissions, Human Review, Model Routing, Privacy, Audit Logs and Software Safety
Open WebUI put coding agents in the model picker. Keep the permission receipt
Open WebUI 0.13.1 can place coding agents beside ordinary chat models and connect them to a real computer through Tether. The useful interface is not merely a model picker. It is a clear receipt showing who asked to touch which tool, file and machine, who approved it and what actually happened.
United States, Collinear AI, Hugging Face, CWE-bench v2, Cybersecurity, Coding Agents, Vulnerability Repair, Common Weakness Enumeration, Open Benchmarks, Held-Out Evaluations, Benchmark Contamination, Harbor, Programmatic Graders, Human Security Review and Responsible AI
CWE-bench v2 opened 100 security fixes. Keep the public and hidden exams separate
The new public set lets anyone inspect the tasks, harness and deterministic graders behind an agent security score. That transparency is useful, but every open exam eventually becomes study material.
United States, Anthropic, Claude Code, Claude Apps Gateway, Coding Agents, Subagents, AI Budgets, Managed Policies, Model Governance, Tool Permissions, Sandboxes, Amazon Bedrock, Guardrails, Audit Evidence and Developer Operations
Claude Code can pool budgets and set subagent effort. Record the policy that actually applied
Claude Code 2.1.316 adds a shared spending pool, a default effort level for subagents and layered gateway policies. The control panel is not the evidence. Keep a per-session receipt showing which settings survived the merge and reached the run.
China, Open-Weight Models, Coding Agents, Multimodal AI, Model Evaluation, Inference Operations, Tool Use and Deployment Governance
DeepSeek made the model faster. The receipt still starts with the harness
DeepSeek says its October 17 update improves reasoning, coding, agents and vision while finishing work with fewer tokens. The useful question is not whether one score went up. It is whether the exact model, encoder, tools and hardware make your work cheaper without making it stranger.
United States, Open-Weight Models, Coding Agents, Model Deployment, Infrastructure, Evaluation, Safety, Licensing and Operational Control
Beam opened the loading bay. Running it is still heavy work
Reflection has released the weights, license, model card, safety report and deployment stack it promised. That makes Beam genuinely inspectable. It does not make a 501-billion-parameter agent model light, cheap or self-governing.
Canada, Coding Agents, Multi-Agent Systems, Software Engineering, Cost Control and Human Oversight
Cohere scaled coding-agent teams. The lead became the bottleneck
Cohere Labs tested who should plan, implement and verify when coding agents work in teams. More workers often helped, but they also multiplied cost and pushed coordination back onto one lead. The useful artifact is not a bigger swarm. It is a ledger showing what each worker discovered and what the lead actually checked.
Czech Republic, Europe, Open Models, Coding Agents, Self-Hosting, Sandboxes and Software Assurance
JetBrains put Mellum2.1 through millions of sandboxes. Open coding models need work logs
JetBrains says its compact open model learned to explore repositories, edit files and check its work across millions of sandbox runs. That is useful training. Production teams still need a reproducible receipt for every command, change, test and permission.
Developer Tools, Coding Agents, Enterprise Migration, Cloud Billing, Security Controls and Operational Continuity
Google stopped selling Gemini Code Assist. The migration plan is the product now
New Standard and Enterprise subscriptions are closed, with license growth and renewals ending on different clocks. Google points customers to Antigravity inside Gemini Enterprise, but a coding assistant and a broader agent platform are not interchangeable line items.
Coding Agents, Extensions, Local Security, Permission Boundaries and Software Supply Chains
Claude Code opened its nervous system to mods, and the plug-in gets your keys
Anthropic's new TypeScript mods can rewrite prompts, intercept tool calls, approve permissions and redraw Claude Code. They can also read files, secrets and sessions with the user's full local authority, so the useful extension layer arrives with a very old security rule: installed code is trusted code.
China, Coding Agents, Context Compression and Model Memory
Peking compressed the coding agent's memory, then paid for every missing detail
Peking University researchers taught coding agents to store older tool results as compact latent tokens. Context shrank and serving sped up, but benchmark success also fell unless more recent text stayed visible.