Benchmarks
2 articles
-
GoogleGemini Optimized ADK Overnight. Your Database Store Missed It.
Google let Gemini 4 Argon run unattended for 11 hours on its 205,000-line Agent Development Kit; two of six proposed optimizations shipped in ADK 2.12.0 (commit 1754b3e, 'Argon generated these changes') and maintainers cut four. Benchmarked against 2.11.0 on an M2 Pro, ADK's pre-model overhead at turn 150 fell from 51 to 25 ms with the in-memory session store and stayed at about 45 ms with the SQLite-backed DatabaseSessionService; before the change the in-memory store was the slower of the two. The LlmAgent import fell from 474 to 357 ms but a fresh process still reached its first model request in 579 and 584 ms, so local evals got faster and the gap to production got wider.
-
OpenAIGPT-5.5 Scored 88.7% on SWE-Bench. That Number Is Misleading.
GPT-5.5 scored 88.7% on SWE-Bench. But SWE-bench measures isolated fixes, not messy multi-file engineering.