3/3 We took 7 weak agents (ranks 7-13, none scoring >45) from the leaderboard & merged them into 1 report/task. Essentially boosting for deep research. The result: New #1 DRB II TotalScore of 64.38. Full write-up here:
@ai21labs
-
How we beat the deep research agent by going the other way
By
–
2/3 Six months ago, the best deep research agent scored ~45. Since, everyone's tried to beat this score with better agents. We went the other way.
-
AI21Labs lands #1 on DeepResearch Bench II without new agents
By
–
1/3 We just landed #1 on DeepResearch Bench II without building a single new agent.
-
Pipeline order is a hyperparameter for optimizing LLM execution strategies
By
–
5/5 Takeaway: pipeline order is a hyperparameter. If you're already paying for parallel rollouts, reuse them – they're relevant context, not just candidate answers. Full write-up: [
https://
ai21.com/blog/first-sca
le-then-enrich-how-the-right-execution-strategy-helped-us-reach-state-of-the-art-on-swe-rebench/?utm_source=org-twitter
…] -
AI21 Labs surpasses Claude Code in efficiency and performance with Test Agent.
By
–
4/5 Still came in ~$0.30 under Claude Code’s spend at a similar score. So we added a lightweight Test Agent that writes repo tests and filters failing patches, pushing our final result to 60.9% – surpassing Claude Code (60.9% vs 56.2%) at the same cost.
-
AI21 Labs: ReAct Agent Performance with Enrichment and Scaling Strategies
By
–
2/5 Started with a baseline: classic ReAct agent (GPT-5.2), single Docker-terminal tool. Baselines on the slice: vanilla 53.8%, enrich-only 55.6%, scale-only (n=5 + LLM judge) 55.4%, enrich-then-scale 57.7%.
-
AI21 Labs: Reversing agent pipeline order achieves SOTA results
By
–
1/5 Our latest Labs in Front piece: Agent pipeline order matters. By reversing a common agent recipe – scale first, enrich second – we reached SOTA on a Dec ‘25 to Mar ‘26 slice (123 issues): 60.9%.
-
AI21 Labs unveils new prompt caching capabilities
By
–
5/5 What this unlocks: – Change one prompt, serve cached results for all other components.
– A/B test any component against identical upstream outputs.
– Run best-of-N inference without N branches collapsing to the same answer. Full design methodology here: -
Designing cache keys to reproduce variance
By
–
2/5 This was tripping up our agent evals. The question we needed to answer: how do you design a cache key that can reproduce variance? Meaning, know when to reuse results and when to recompute.
-
Inefficient caching for LLMs with parallel strategies
By
–
1/5 Caching relies on a deterministic assumption: same input, same output. This breaks down with LLMs, particularly with agents using parallel rollout strategies like best-of-N sampling.