Aziz Latipov

Agents with Tools Fail Together

One public asset for the axis the earlier notes set aside: 175 software agents, each a model wrapped in its own tools and workflow, scored on the same 500 SWE-bench Verified bugs. The seven best carry 1.4 effective independent votes, not seven; a majority of them resolves 79.4% o...

Agents Break Together on Misuse, Apart on Injection

Do safety breaches of computer-use agents co-occur across model families the way correctness errors do? Five agents from four vendors on 150 OS-Harm tasks, pre-registered, zero cost. On deliberate misuse the agents comply together (tetrachoric ρ 0.67, 1.4 effective votes of 5); o...

Where the Swarm Pays

A ledger of what multiplying LLM agents buys on 240 hard problems across seven reasoning models from five families, revised three times: after a label audit (on the problems where all seven models were scored wrong, the gold label, the question or the grader was at fault on 45 of...

Debate Is Not Sampling

The modal ceiling of test-time scaling does not transfer to multi-round debate: on 9,600 public traces three-agent debate beats the round-0 modal-hit rate on weak models and corrects 31–45% of problems where all agents started wrong. The correlation ceiling does transfer: inter-a...

Debate Is Not Sampling

The modal ceiling of test-time scaling does not transfer to multi-round debate: on 9,600 public traces three-agent debate beats the round-0 modal-hit rate on weak models and corrects 31–45% of problems where all agents started wrong. The correlation ceiling does transfer: inter-a...