$ cat briefs/daily/2026-11-04.md
2026-11-04
November 4, 2026 (Wed)
2 top stories · 1 additional paper pick · 1 tooling update
Top Stories
1. Antigravity puts chemistry calculations behind an agent workflow
Google's 2026-11-03 walkthrough combines literature search, compound curation, matched molecular pair analysis and proposed synthesis routes. The model orchestrates; deterministic RDKit tools perform molecular calculations. Google reports an hour for the demonstrated workflow, while explicitly requiring synthesis and experimental validation of proposed compounds.
Why it matters: domain tools make the calculation step inspectable, but faster hypothesis generation is not validated drug discovery. This is a vendor demonstration. → Google DeepMind (source)
2. LeanGuard checks the call before the tool executes
A Lean4 policy check fails closed and can return feedback or a proved repair. Authors report direct prompt-injection success falling from 57.9% to 5.9% on DTap. Benign success is unchanged in the telecom setting, where benign tasks need no writes; that qualification matters.
Why it matters: action checks provide an explicit enforcement point, but correctness against an encoded policy does not establish policy completeness or general safety. → LeanGuard — checking policy before tool execution (source)
Paper Picks
SWE-Trail: use previous trajectories to choose where to collect more experience, without inserting that history into the coding agent's context. A 124B student improves over independent sampling by 2.33 percentage points on SWE-bench Multilingual and 1.60 percentage points on SWE-bench Pro. Read it for the history-free ablation: more solved collection trajectories alone barely improved Multilingual. Total collection cost is not quantified in the abstract. → SWE-Trail — using history without teaching dependence on it (source)
Watch
- LLM 0.37: the maintainer adds
llm openai endpoint addfor compatible endpoints and content-type support for mixed binary embedding inputs. Useful interface changes; backend behavior was not independently tested here. → Agents (LLM Agents) (source)
New in Wiki
- LeanGuard — checking policy before tool execution — methods, results and policy-coverage limits. (source)
- SWE-Trail — using history without teaching dependence on it — collection procedure, student evaluation and open cost questions. (source)
Updates
- Agent Runtime Containment, Agentic Reinforcement Learning and Agents (LLM Agents) connect today's papers to existing research. (guardrails) (training)