Deep research agents get measured: LifeSearch, PANINI-DR and a fast-and-slow thinker

The benchmark that embarrasses everyone

Deep research agents have been graded mostly on academic trivia; LifeSearch grades them on life. In a preprint, the authors describe 203 tasks across seven domains, built from real-user query seeds and backed by over 4,500 hours of human annotation and 4,000 fine-grained rubrics. Twelve advanced agents took the test. The best, GPT-5.6 Sol, managed 63.4%. The authors report the scores track human judgment closely (Spearman's rho 0.89, Kendall's tau 0.81), so the low numbers are a property of the agents, not the test. m:…2254

For developers shipping research agents, 63.4% on real-world tasks is the number to internalize. It says the gap between 'impressive demo' and 'reliable coworker' is still roughly a third of the work — and that any vendor claiming production-readiness should be asked for their LifeSearch score, or an equivalent, before the contract. m:…2254

In a preprint, the authors note the benchmark spans 203 Chinese and English tasks — bilingual from the start — and argue that deep research agents are currently insufficiently evaluated in real-world life and productivity settings. That is the gap LifeSearch is meant to fill: not harder trivia, but the messy questions people actually ask. m:…2254

A workspace, not a longer prompt

PANINI-DR's answer to the reliability problem is architectural: stop asking one big model call to do everything. In a preprint, the authors describe a multi-agent harness that splits the job — one layer coordinates the overall investigation, another verifies individual subquestions — with both sharing a Generative Semantic Workspace that records subquestions, dependencies and evidence across calls. The headline they report: 73.3% accuracy on BrowseComp-Plus, 16.0 percentage points above the next-best harness, and the strongest average across three benchmarks. m:…4630

The mechanism the authors credit is check-before-propagate: findings get verified before errors cascade into later steps, and the shared workspace makes answers recoverable from saved evidence. If you build research agents, the design lesson is that persistent, evidence-backed state beats cleverer prompting — structure the loop so mistakes get caught, not just made faster. m:…4630

The evaluation is broader than BrowseComp-Plus alone. In a preprint, the authors report testing on FRAMES and MoNaCo as well, with PANINI-DR taking the strongest average across all three, and note it correctly answered 26 of 50 hard questions on BrowseComp-Plus. A separate verification agent checks proposed answers against subquestions and evidence, and the authors report the highest answer recoverability from saved evidence on all three benchmarks — meaning when it fails, you can at least see why. m:…4630

Fast and slow, on purpose

JevGPT uses a dual-system architecture where System Two handles deliberative planning — setting evidence objectives, integrating findings — while System One does fast relevance assessment and source inspection in parallel. In a preprint, the authors report leading task performance on fixed-corpus question answering and document ranking, with faster completion and lower cost than comparably effective research agents, plus broader source coverage in the research loop. m:…3639

Read the three together and the week's message for builders is coherent: measure on real tasks (LifeSearch), structure the investigation so errors get caught (PANINI-DR), and split thinking into fast and slow lanes (JevGPT). All three are preprints, so treat the numbers as the authors' claims — but the design patterns are cheap to steal, and that is what matters most. m:…2254 m:…4630 m:…3639

The division of labor is the detail to copy. In a preprint, the authors describe System Two setting evidence objectives and integrating findings while System One locates support, assesses relevance and guides further inspection in parallel. Their evidence analysis reports broader source inspection and greater coverage of supporting documents in the research loop. Slow thinking plans the investigation; fast thinking works the sources — and the parallelism is where the cost savings come from. m:…3639

What the ablations say to steal

If you only take one implementation idea from the week, take PANINI-DR's ablation result. In a preprint, the authors report that structured planning, persistent state and evidence-based verification each proved critical to the performance gains — remove any one and the numbers drop. That is a checklist for any research-agent builder: a plan the agent can inspect, state that survives across model calls, and a verification step before findings propagate. Most production research agents I have seen described have at most one of the three. m:…4630

And if you only take one economic idea, take JevGPT's: parallelism in the fast lane. In a preprint, the authors run independent source judgments in parallel while the slow lane integrates — which is why the system reports both better coverage and lower cost. Research agents spend most of their wall-clock waiting on sources; parallelizing the judgment step is the closest thing to free lunch in this design space. m:…3639