In September 2026 I shared a series of LinkedIn‑style notes about three emerging AI research themes: AutoGym for automated RL environment generation, bias and cost‑inflation in personal agents that access private data, and the Agensh multi‑agent coding harness. I also compared two System One decision‑making models, Jev and the Contrastive Language Model (CLM).
- AutoGym (Amazon AGI) generates an RL task, environment, and verifier from a minimal seed.
I decided to adopt AutoGym for rapid prototyping of RL environments, reducing manual engineering overhead.
I will implement bias‑mitigation safeguards in personal agents: limit memory/inbox access, anonymize financial cues, and evaluate cost recommendations across model sizes.
I will leverage the Agensh decentralized multi‑agent harness when scaling coding workloads, as larger teams raise pass rates.
I prefer CLM over Jev for verification‑heavy pipelines where latency matters, especially for long‑horizon tasks.
The study ran 325 K experiments on 13 models across flights, health insurance and graduate programs.
The gaps reach $198 per flight and $284 per month for insurance with Claude Opus 4.8.
On the five hardest ProgramBench tasks with GPT-5.6-sol, increasing agents from 1 to 128 raises the mean final test-pass rate from 19.31% to 28.78%.
On pandoc, 1,024 agents take the test-pass rate from 33.89% to 55.06%.
CLM is 9x faster than Jev.
- Personal‑Agent Bias Study uncovered cost gaps up to $198 per flight and $284 per month insurance.
- Agensh (Microsoft Research) uses a decentralized harness of >1 K agents that improves pass rates as team size grows.
- System One models: CLM is ~9× faster than Jev and verifies better on long‑horizon tasks.