Mercor, a profitable Series C AI data company valued at $10 billion, is hiring a Member of Technical Staff for its Enterprise Evals Platform in San Francisco, offering a salary of $220,000–$425,000 pe…
Mupt AI released self-bench, an open-source tool that automatically converts private, completed pull requests into held-out coding-agent evals, generating validated Harbor datasets with a single CLI c…
JetBrains' PyCharm introduced the Agent Environment Coordinator skill, which provides AI agents with the correct Python interpreter information, boosting task success rates from 68% to 98% on average …
Antigma Labs released Ante, a self-contained coding agent in a single ~15MB Rust binary that runs offline with zero runtime dependencies, achieving 82.7% on Terminal-Bench 2.1 (368/445 trials) using D…
A benchmark of 2,040 CRM agent rollouts found that over half of the small open-weights model gemma-4-E4B-it's 'failures' actually contained the correct answer but were graded zero because the model ne…
BoundaryBench, a new open-source benchmark from the developer community, measures how much capability coding agents lose when running inside enterprise/NIST-derived hardened sandboxes, with live resul…
A controlled test of the token-saving tool Headroom on 25 senior software engineering tasks using Anthropic's Claude found that while it cut input tokens by 39%, it did not reduce costs due to cache b…
JetBrains tested the Caveman token-compression skill on Claude Code 2.1.200 and found it saves only about 8.5% of output tokens on real agentic tasks, far below the advertised 65%. The skill, which fo…
Letta AI introduced the Trajectory package, a standardized data format for agent experience that reduces token counts by up to 5.6× compared to native session formats, enabling agents to learn across …
The team behind Terminal-Bench and Harbor released Frontier-Bench v0.1, a continuous benchmark with 74 tasks across 7 domains that measures agent abilities at the frontier, with the best models achiev…
Letta released the open-source trajectory package on July 24, 2026, a standard format for recording coding-agent sessions across harnesses including Claude Code, Codex, and Letta Code. The package red…
A developer built nano-harness, a coding agent in ~970 lines of Python that scored 59.6% (53/89) on the full Terminal-Bench 2.0 suite using Claude Opus 4.8. The harness uses only three tools—bash, rea…
IntelligenceLab requests Hugging Face to add Long-Horizon-Terminal-Bench (LHTB) to the Benchmark allow-list. LHTB is a 46-task benchmark that tests LLM agents on long-horizon terminal tasks using hidd…
JetBrains AI tested the rtk (Rust Token Killer) CLI proxy, which claims to cut agent tokens by 60–90%, and found that most agent bytes never touch the hook, with measured token savings far below adver…
The Cloud Native Computing Foundation (CNCF) announced on July 16 that Broadcom has elevated its membership to Platinum, following Broadcom's contribution of the Velero backup and migration tool to th…
Harbor, a new cross-platform note-taking app founded by Spicer Matthews, promises a private, encrypted second brain that users own for life, with a three-year price lock and a cap on annual increases,…
A new open-source harness from Tensorlake demonstrates that agent evaluations are vulnerable to cheating unless per-task isolation is enforced, showing a 53% lie rate when agents can tamper with test …
Prime Intellect released verifiers v1 (version 0.2.0), a rewritten environment stack for agentic reinforcement learning and evaluations that decouples tasksets, harnesses, and runtimes into composable…
Google has updated its Android Bench ranking system for AI models, switching to the Harbor framework and opening submissions to developers. Anthropic's Claude Fable 5 tops the new list with a score of…
Google has updated its Android Bench LLM leaderboard with the Harbor framework, adding new models such as Claude Fable 5 and GPT 5.5, and inviting community contributions to the benchmark. The changes…