AI Builder Brief · 2026.11.06
AI Builder Brief · November 6, 2026
AI-assisted research may be accelerating the frontier, but the most useful engineering lessons concern who controls agent actions, how changes are undone, and what successful work actually costs.
METR estimates Anthropic’s capability progress has accelerated roughly twofold
After eleven weeks of access, METR estimates Anthropic’s benchmark-progress rate increased 1.9× since October 2025, with a 95% interval of 1.5–2.8×, and judges AI assistance probably explains most current acceleration. It distinguishes faster research from faster compute scaling and from staff’s much larger self-reported productivity gains; these quantities are not interchangeable.
Dropbox explains host-enforced permissions and undo for file-changing agents
Dropbox’s newly published implementation puts credentials and permission checks outside each agent VM, intercepting both tool calls and ordinary programs; a Pi adapter shares the interface’s Codex protocol without rebuilding those controls. File versions are attributed to conversations, so undo restores only their changes and creates a separate copy when another person has subsequently edited an affected file.
Claude Code publishes a narrow guard after renewed email-leak reports
A new issue reports the agent inserting the account email into outbound User-Agent headers; a contributor says a default-on privacy mod exists in 2.1.317 and later. The published implementation refuses certain shell calls containing both that address and User-Agent indicators, but explicitly leaves web/MCP tools, computed values and other transmission shapes outside its coverage.
Thinking Machines launches Inkling 2 Small, with weights still pending
The 594B-total, 17B-active model is available for sampling and fine-tuning through Tinker; Apache 2.0 weights are intended for next week after external safety testing. Its report describes native NVFP4 pretraining and a unified nine-day RL run exceeding 17 million rollouts, including long tasks with compaction, rather than claiming the whole model trained in 68 hours.
NVIDIA turns recorded agent failures into tests and targeted harness fixes
The newly published explanation connects trace IDs and an explicit behavior contract to generated Harbor tests, a single instruction change, and development versus hidden holdout comparisons. In its fictional enterprise assistant, unwanted message-sending falls enough to improve development passes from 3/6 to 6/6 and holdouts from 4/6 to 6/6; linked artifacts expose the exact candidate and verifier.
TCSExam separates complete proofs from useful but unfinished research progress
Moonshot and Tsinghua researchers evaluate agents on 100 tasks spanning Lean proofs, solver engineering, constructions and open conjectures, with a 20-task Lite suite. The important mechanism is preserving distinct verification and scoring rules: intermediate lemmas are not complete proofs, solver scores use frozen references, and improvements over a task reference require separate novelty checks.
Cursor adds app hosting, databases and explicit publication to cloud agents
Paid plans with cloud agents enabled can now build and deploy applications with Postgres, Cursor SSO and hosting, from new or existing Cursor Origin/GitHub repositories. Deterministic and agentic security checks precede an explicit publish action; team admins can inspect apps and secrets, add review rules and unpublish applications, while individual plans can share publicly.
Baidu releases ERNIE’s pretraining checkpoint for custom post-training and agents
ERNIE-5.5-Flash-Base exposes FP8 and BF16 weights for a 280B-total, 14B-active model trained on 33.5T tokens, before mid- or post-training. Its architecture combines heavily compressed attention with periodic finer-grained access, multiple expert routers, and reuse of the final layer for future-token prediction; the card provides specific SGLang and vLLM-plugin recipes.
Conductor joins Meta and gives cloud users a near-term migration deadline
The agent-workspace product, originally launched July 2025, stops new signups, makes Conductor Cloud read-only November 13, and shuts it down November 20, when cloud-data deletion begins. Existing users can transfer individual workspaces and chats to their Mac or export all data; local Conductor receives models and critical fixes until early 2027.
Poki’s first users expose readback costs and drive an iPhone-control fix
A user’s November 6 physical-iPhone trace attributes most of a seven-second action to readback and repeated process startup, rather than model inference. The maintainer’s 15:33 UTC change keeps a helper and input session resident; five-run Wi-Fi comparisons reduce a six-tap calculator sequence from 51.7 to 37.2 seconds, while verification remains the default.
← All AI Builder Brief entries ← Older (Nov 5) Source repository ↗