Find published research by company, product, platform or technology.
Showing 1–20 of 122 dossiers
The pilot attacks a persistent evaluation trade-off: labs do not want to reveal frontier-model internals, while evaluators do not want benchmark prompts leaking back to the model provider. DeepMind says a Singapore AI Safety Institute pilot kept both sides’ sensitive assets hidden during execution.
The important development is not simply another AI security mishap. Anthropic found a fourth incident missed by its first review, widened the search to hundreds of millions of transcripts, revised its causal interpretation and invited an external evaluator to investigate the full record.
Railway Cloud Agents are managed, persistent development machines rather than a new model or harness. They reuse developers’ existing agent credentials, sleep when disconnected by default, retain disk state, and live inside Railway project environments—blurring the boundary between remote coding workspace and deployment platform.
The useful part of Smaug Agentic is not another frontier-style benchmark claim. Abacus.AI is publishing a drop-in Kimi K3 derivative that targets a specific production failure mode in coding agents: long runs that burn the reasoning budget without converging. The weights and model card are public, but the training data is not disclosed and the benchmark gains remain vendor-produced.
The workflow shift is continuity rather than another model upgrade: one Kiro agent session can outlive the laptop that started it. Cloud configuration can also carry agent setup across environments, although enterprise governance is not identical between local and web/cloud surfaces.
GitHub Copilot can now turn Slack or Teams threads into collaborative cloud-agent sessions. Teammates can add context and steer the work in public, while repository permissions, agent budgets and optional extra PR approvals remain the main control boundaries.
Qwen3.8-27B is now available as open weights on Hugging Face and ModelScope. For builders, the important change is not another benchmark bump: a comparatively compact 27B model now combines native vision, long context, controllable reasoning and OpenAI-compatible serving paths for local or self-hosted coding and agent workloads.
Funes treats agent memory as user-owned data rather than a hosted account feature: retrieval and reranking run locally, provenance stays attached to recalled passages, and cross-machine sharing is optional. The main risk is that publishing session-derived memory can still expose secrets if redaction misses them.
Google appears to have completed a talent-focused Mechanize deal: the startup still exists, but much of the team that builds coding-agent training environments and evaluations has moved into Google’s model-development work.
Hy4 preview is a very large sparse model with public full and FP8 weights, native speculative decoding and a 1M-token context path. Its open release makes Tencent’s claims testable, while the 1.56TB full checkpoint keeps self-hosting firmly in server-scale territory.
GLM-5.3-Flash combines open weights, multimodal coding/agent capability and an 18B-active sparse architecture with a large anonymous pre-launch trial. Z.ai has already issued a chat-template correction for early downloads, showing that day-one self-hosted deployments need artifact-level validation as well as model benchmarking.
Microsoft's Agent Host already detached coding-agent sessions from one editor window. Agent Merge shows what that architecture enables: a long-running worker can keep cycling through PR feedback and CI state rather than stopping after one code-generation turn. The feature is still Preview and needs the same review, permission and side-effect controls as any autonomous delivery loop.
OpenAI’s internal data turns “agents make researchers faster” into a measurable operating model: heavy concurrent agent use, record experiment throughput and rising task complexity, alongside high token spend and persistent human intervention on longer work.
Gemini 3.8 Flash keeps 3.7 Flash’s promotional per-token rate and Flash-tier latency, but early independent analysis suggests harder reasoning can increase tokens consumed per task. A separate 3.8 Flash Cyber model is available only through Google’s Fairwind defensive-security program.
Data Agent Kit turns Google Cloud’s data tooling into an agent-callable developer surface. The useful shift is portability across coding assistants, but the kit remains an open-source integration layer around Google Cloud services rather than a vendor-neutral data runtime.
WebKit’s Safari MCP server turns browser debugging into an agent-callable interface. It runs locally and makes no network calls itself, but captured page data is sent directly to the connected agent, so browser-session trust and model data handling become part of the development security model.
The staged release is complete: GLM-5.3’s public weights and serving artifacts are now available. That makes Z.ai’s coding and cyber-capability claims independently testable while turning the earlier safety delay into a concrete self-hosting and audit decision.
The important shift is that agent orchestration itself becomes a managed API surface: context compaction, tool discovery, programmatic tool calls and subagent coordination can now come from OpenAI’s maintained Codex harness rather than an application team rebuilding those layers.
Agent Identity is moving from a standalone credential boundary into a mainstream serverless runtime. Cloud Run can now assign agent identities and register agents/MCP servers automatically, reducing custom discovery and identity plumbing while keeping the runtime integration itself in Preview.
Grafana’s GA agent-observability stack can track latency, tokens, cost and conversations, score live traffic with deterministic or LLM-based evaluators, route failures into test collections, compare experiments and use those results as pull-request gates. Evaluator quality and instrumentation coverage remain the main limits.