Glean 拾遗
中 EN
Submit a link
Daily · timeline

A few picks a day.

Mon, Aug 17, 2026 3picks
← 08-16
Calendar ▾
2026 · 08
MoTuWeThFrSaSu ·····12345678910111213141516171819202122232425262728293031
has picks today
06:00

Maximizing the value of your Claude Code sessions

把 Claude Code 的每一分钱花在刀刃上：token、缓存与会话管理

Claude Code bills tokens in two phases: prefill reads the whole request, decode generates output one token at a time, which is why output costs about 5x input. Prompt caching is automatic: shared prefixes are served at 0.1x, but switching /model, /effort, or fast mode, or running /compact, invalidates the cache and forces a full-price re-prefill. Session cost is determined by how many tokens enter the context, how many turns they stay, and how many contexts run in parallel. The guide offers concrete tactics: use @-mentions to avoid Read calls, put quiet flags for daily commands into CLAUDE.md, spill outputs over 30,000 characters to a file, use /rewind instead of /compact to trim dead ends, and offload noisy jobs to subagents (optionally pinned to haiku). Cache expiry is one hour on subscription and five minutes on API keys, extendable via ENABLE_PROMPT_CACHING_1H=1. Practical for heavy Claude Code users.

claude.com · 13 min · Claude Code · Context Engineering · Cost Optimization · Prompt Caching · Token-Optimization
06:00

Graph Engineering: from 1 prompt to 100 agents in one system

删掉不存在的箭头：从 1 条 prompt 到 100 个 agent 的图工程八步

This article argues that the workflow itself is the engineering target: alongside prompt, context, harness, and loop engineering, there is a fifth layer the author calls graph engineering. The core observation is that default linear agent flows confuse sequence with dependency: most arrows do not read an upstream result, so cutting them is what enables real parallelism. A hundred agents, in this view, are not a hundred roles but one role instantiated a hundred times with its own slice of the problem. The eight-step method covers node contracts, four topologies (chain, fan, router, controlled cycle), when joins are worth waiting for, separating probabilistic classification from deterministic routing, placing independent verifier nodes, and keeping durable state. It explicitly warns that parallel breadth is expensive, citing Anthropic's multi-agent research system consuming roughly 15x the tokens of a normal chat. Useful for engineers moving from a single prompt to a multi-agent production system, though there is no runnable code.

x.com · 10 min · Agent Architecture · Agents · Context Engineering · Cost Optimization · Orchestration
06:00

Your Agent Remembers Everything and Understands Nothing

记忆不等于理解：Agent 缺少的跨会话归纳层

The article argues that agent memory systems today are stuck at the storage-and-recall stage: they return the right facts but miss the structural patterns spanning multiple conversations. Using a project-management assistant as a running example, it shows three engineers filing blockers that are individually accurate yet actually point to one shared root cause — a delayed auth-service refactor. Better retrieval, rerankers, or wider search scopes don’t surface that chain because the insight was never stored as a discrete piece of content. The proposed mechanism is Zep’s Observations: a deterministic algorithm first reduces every fact to a signature of two entities plus the relationship type, then builds an episode graph where conversations are nodes and shared signatures are edges, producing connected components that expose transitive dependency chains. An LLM then writes a name and summary, but grouping is never left to the model. No embeddings, no semantic similarity — pure graph topology. Relevant for engineers working on agent memory, knowledge graphs, and context engineering. Note: this is a sponsored, technically dense explainer for Zep.

x.com · 13 min · Ai-Memory · Context Engineering · Knowledge Graph