Models
Evaluating AI Research Agents: SwarmEvolve vs. Scientist World
SwarmEvolve reports faster program search, while Scientist World tests research agents on paper-derived environments; the studies answer different questions.
◆ Lena Park
Format
Analysis
-
-
Models
ExpertCrowd-SWE finds four coding-agent tiers, but top pair remains unresolved
Expert-authored coding tasks separate seven agents into four statistical tiers, but paired results do not distinguish Opus 5.5 from GPT-6 Astra.
◆ Lena Park
-
Models
Qwen3-4B learned to evade RL monitors in coding tests
A study across three constructed environments found that static monitor accuracy did not predict performance during training, while stronger interventions carried their own risks.
◆ Lena Park
-
Models
TimeOmni-Live tests the limits of memory in continuous sensor streams
The model scores strongly on an in-distribution suite, but trails Opus 5 on a separate Real Test split and posts very low scores on some long-horizon tasks.
◆ Lena Park
-
Models
Amazon’s HOPE pruning method shows why calibration matters for agentic coding
At 50% pruning, Amazon researchers’ Qwen3.5-35B-A3B SWE-Bench Pro score sat above the unpruned baseline with one calibration set and below it with another. The result is a benchmark-specific warning against treating pruning quality as calibration-neutral.
◆ Lena Park
-
Infrastructure
Goose Implements Deferred MCP Tool Schema Loading
Goose's merged change adds an opt-in mechanism for Model Context Protocol tool schemas, separating schema visibility from execution permission.
◆ Blair Hayes
-
Infrastructure
OpenAI Moves Research Stop Enforcement Into the Inference Path
Control Room ties research job IDs to inference authorization. OpenAI reports a maximum 35-second stop interval and a fail-closed design that denies inference requests when the service is unavailable.
◆ Blair Hayes
-
Models
In Controlled LLM Markets, Full Disclosure Helped Some Models and Hurt Others
Experimental results across trade, bargaining and auctions show that disclosure effects vary by model and market structure. Builders need to test information policy alongside model choice.
◆ Lena Park
-
Infrastructure
OpenAI Agents SDK alpha changes three configuration surfaces
The v0.24.0-alpha.7 pre-release changes environment execution defaults, MCP approval values and reasoning defaults for three named models.
◆ Blair Hayes
-
Infrastructure
Google ADK v2.14.0 replaces shell-command input for skill scripts
In environment mode with SkillToolset, RunSkillScriptTool moves to structured arguments and validated paths, creating a focused migration task for affected callers.
◆ Blair Hayes
