MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory — a new benchmark of 155 multi-session trajectories across four domains finds every one of five tested agent-memory systems is vulnerable to a malicious instruction planted in one session and later acted on, with how much it matters varying by system design.
The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks — pairing frontier coding agents with reusable “agent-native primitives” (HEELO) raises Terminal-Bench 4.0 resolution from 37.3% to 55.5% on the same backbone, and on 46 real-world tasks takes Claude Opus 5.5 with Claude Code from 9 solved tasks alone to 39 when connected via HEELO’s MCP server, at 1.9x less time and 1.55x lower cost per completed task.
How Much Conversation History Does an AI Agent Need? — testing fixed context budgets on 1,100 questions, the authors find the score gain per added token shrinks at the largest budgets tested, and which message-retention strategy best matches the full conversation depends on the model and on whether the answer needs an exact earlier reference (unverified: the paper reports no specific score values for this).
Hide your slop PR descriptions behind <details><summary> — Simon Willison argues PRs are his real unit of work, and since coding agents already write usable PR descriptions, the fix for noisy ones is collapsing them behind a <details><summary> toggle rather than discarding them.
