Issue #3

Anthropic lays out how it designs memory for Claude: notes written during a session, plus “dreaming”, an offline pass that rereads past transcripts to correct notes, build linked wikis, and decide which notes get loaded into every new session. Newer models judge what’s worth saving much better. Playing Pokémon, Sonnet 3.7 wrote 413 files and earned one badge, while Fable 5 wrote about ten and earned four. Telling Opus models to “record only failures” cost them much of what memory gained them. On a 420M-token company corpus, a wiki built over 128 dreaming sessions lifted scores from 13% to 48%, and wiki plus search beat search alone (61% vs. 53%) while saving about 20,000 tokens per question. In a legal task, giving each contract its own page written by a single agent, instead of a shared table, raised the facts kept from 41 of 57 to all 57.

Transluce trained “oversight” models to predict how a small model (Qwen3.5-2B) would answer prompts after being fine-tuned on a given bundle of documents. It used up to 328,000 examples and oversight models from 0.8B to 14B parameters. Prediction quality followed a Chinchilla-style power law in model size and data, with optimal data growing faster than optimal model size, but it flattened between 4B and 9B and hit an apparent floor that more compute doesn’t remove. The oversight model trained from the same 2B base as the subject beat larger ones, which the authors call a privileged-access effect. On harder tests (datasets that look benign but instill behaviors such as recommending drunk driving), the 2B and 9B oversight models beat the pre-finetuning baseline, though out-of-distribution loss fell more slowly than validation loss.

Google traced coding-agent sessions across thousands of its engineers, using inline TODO(bug/...) comments as a measurable signal of stale context. 13% of sessions came across one. In 6% of those encounters the agent acted on the bug, and 83% of those actions were unprompted. Of the unprompted detours, 41% chased bugs that were already fixed or abandoned, and 57% expanded scope into open bugs nobody asked about. Controlling for developer, task and model, these detours added 70% more steps, 77% more tokens and 105% more wait time per session. Google then had agents dig through bug and file history to confirm which TODOs were obsolete and propose deleting them. Over six months they removed thousands, and reviewers accepted more than 95% of the changes.

Anthropic classified about 207,000 research sessions on Claude and Claude Code, using only privacy-preserving labels, and paired them with a survey of 528 scientists. Data analysis accounts for about a third of Claude sessions and nearly half of Claude Code sessions. Ideation and data gathering, which scientists rate most important, get little use. Tasks in the top fifth on an AI capability score get two to four times the use their share of researcher time would predict, and that score correlates strongly with how suitable a task is for reinforcement learning (0.78). With wide adoption of today’s capabilities, the model has AI taking on 27–52% of researcher time and raising output by 1.35x to about 2x. Even automating all non-physical work caps lab sciences at 3–4x, because physical steps take about 40% of researcher time there.

LangChain argues that a standard ReAct loop can’t be responsive and productive at the same time, because each long tool call blocks it. The fix is to split the agent into an orchestrator that stays with the user and async subagents that run in the background. The orchestrator manages them with five tools: start, check, update, cancel and list. A follow-up instruction either interrupts and restarts a subagent (for corrections) or waits until its current run finishes (for additions). Every unfinished task is listed in the orchestrator’s system prompt. Results arrive mid-run if the orchestrator is busy, wake it if it’s idle, and wait if it’s paused on an approval. The post points to the same pattern in Claude Code, OpenAI’s dots and DeepSeek Harness. Task notifications are still in beta.

The post sorts agent retrieval failures into three kinds: inefficient searching, too much context at once, and irrelevant or stale context. It walks through fixes using a weekly competitive-analysis agent as the example. Search names and keywords before reading files (glob and grep, then page through results 100 lines at a time), add semantic search only when the wording is unpredictable, and let the agent decide when to search again instead of retrieving once up front. Results over about 20,000 tokens are saved to a file, and the agent gets back only the path and a preview. For custom retrieval tools, the pattern is “retrieve, offload, delegate”: write results to files, return only the paths, and have subagents read each file and report back only the key facts and the source. To stop stale context, carry dates through into the text the model sees, and treat retrieved text as untrusted.