01Agents
Tools
Agents and dev tools worth using, the releases that change the work, and status-page incidents worth watching.
Other
Negative Self-Distillation: Learning to Reason by Avoiding Flaws
Label-free self-distillation that pushes models away from their own flawed reasoning traces — no teacher model or external labels. Worth a look if you're tuning reasoning models that drive agent tool-calling.
HarvestBench: Measuring Whether LLM Agents Will Pay to Avoid Killing Animals
Benchmark probing whether LLM agents will pay to avoid killing animals in economic tradeoffs — measuring value alignment, not task success. Useful signal if you're evaluating agent decision-making under real-world constraints.
OpenDiscoveryTrace: Process Traces for Evaluating AI Scientist Workflows
OpenDiscoveryTrace — process traces for evaluating AI scientist workflows. If you run autonomous research agents, this gives you a way to measure process quality (what the agent did), not just final-answer accuracy.
How a researcher uses Codex and ChatGPT to search for new antimicrobial molecules
Case study: a researcher runs Codex + ChatGPT to search for novel antimicrobial molecules — coding agents doing scientific discovery, not just writing software. Useful signal on where agent capability is actually landing outside the dev-tool bubble.
An Autonomous GeoAI Agent for Arctic Eco-Navigation
An autonomous GeoAI agent for Arctic eco-navigation — domain-specific agent application combining geospatial reasoning with autonomous decision-making. Niche, but a concrete example of agents moving into physical-world domains.
PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents
PARSER splits long-document QA into parallel chunk reads plus iterative scatter-gather reasoning — targets the accuracy cliff agents hit on multi-hop questions past ~100k tokens. Directly relevant if your agents do retrieval over long docs.
Introducing the Agents API
OpenAI's Agents API — hosted agent runtime with tool use, sub-agent handoffs, and built-in tracing. The signal: orchestration moves from your code into the API, and OpenAI wants to own that layer.
Do Agents Know When They Succeed? Calibrating Agent Confidence from Internal Representations
Calibrates agent confidence from internal representations rather than final outputs — a step toward agents that know when they've actually succeeded and when to ask for help.
Subagents vs Agent Skills: Executing Reusable Knowledge for Long-Horizon Agentic Tasks
Head-to-head of subagents versus reusable 'skills' for long-horizon tasks — evidence on when spawning a subagent beats loading procedural knowledge, and vice versa.
AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
AgentGrad optimizes multi-agent prompts by sequentially intervening to identify which agent is failing, then clustering textual gradients to target fixes — a more surgical alternative to blanket prompt tuning across the whole system.