Glean 拾遗
Daily · timeline

A few picks a day.

Tue, Aug 18, 2026 2picks
← 08-17
Calendar ▾
2026 · 08
MoTuWeThFrSaSu ·····12345678910111213141516171819202122232425262728293031
has picks today
06:00

Graph Engineering explained: what it is, when to use, when not to

图工程入门:为 AI 工作流建模,该并行时就并行

This article frames multi-agent workflow design as 'graph engineering': nodes do bounded jobs with defined contracts, edges carry real outputs, and a 'fake-edge test' exposes unnecessary waits in linear chains. The core pattern is the diamond—fan out parallel research, reduce with plain code, synthesize with one final agent—plus a strict rule that verifiers must run on fresh context, never the worker's chat. It includes copy-paste specs for Claude Code dynamic workflows (triggered by the word 'workflow') covering research, SEO drafting, go-to-market, refactoring, and discovery loops, along with three failure modes: context collapse, false independence, and silent node failure. It also cites honest economics: the Bun rewrite example ran ~50 workflows with up to 64 agents over 11 days, costing roughly $165,000 and requiring heavy human supervision. For engineers already using Claude Code and looking to move from sequential prompts to parallel orchestration, this is a useful but beginner-leaning guide.

06:00

Harness Swap Lifts Same Model from 46.7% to 66.7% Success

同一模型换 Harness 成功率从 46.7% 升至 66.7%,缓存命中率 99.93% 的成本账

In a Composio benchmark, DeepSeek V4 Flash running across 8 agent harnesses passed 20 of 30 hard tasks with Pi (66.7%), while Claude Code, Codex, and Deep Agents each passed 16. Pi's cost per successful task was $0.028, about 1/7 of Claude Code's $0.195. Community numbers show Pi and DeepSeek hitting 99.93% cache hits, processing nearly 1B input tokens for just $2.65 instead of an estimated $132. The article explains DeepSeek's prefix cache: matches must start from the first token, so harnesses must keep the prompt head stable. Reasonix and pi-deepseek-cache demonstrate concrete tactics: frozen environment summaries at startup, append-only context, separate sessions for executor and planner, and hash-cached deterministic summaries, cutting input token prices by 98-99%. Useful for engineers choosing agent harnesses or optimizing LLM API spend.