Glean 拾遗
Recent picks

6picks · chronological

11-13

Graph Engineering: Building a CI Triage Agent Graph

A how-to for replacing a single agent loop with a graph of specialists: nodes each own one verb and a scoped context, edges are deterministic routing code, a shared state file carries history across runs, and a gate decides done or not done. The worked example is a GitHub Actions failure triage system, with copy-ready agent definitions for a classifier and a bug fixer, edge rules that keep hard constraints out of prompts, a STATE.md layout, and a test-gate prompt. Autonomy is staged from manual /goal to scheduled /loop to cloud Routines. The author concedes most teams should not build one yet, and names comprehension debt as the cost of a graph that ships code nobody reads.

11-11

Self-hosted workspace that assigns issues to AI coding agents

Multica is a source-available, self-hostable workspace that treats AI coding agents as team members. Instead of juggling separate terminal sessions, you assign an issue to an agent; it picks up the work on a runtime you control, comments as it goes, and hands back a diff for review. It drives 29 existing agent CLIs—Claude Code, Codex, Cursor, Copilot, Kimi, and others—rather than shipping its own model, so changing providers is a dropdown. The stack is a Go backend with PostgreSQL, a Next.js web app, Electron desktop, and Expo mobile clients. Execution logs, mid-run steering, usage analytics, review gates, and self-hosted Git integration keep humans in the loop. It targets engineering teams already running multiple coding agents who want shared assignment, audit, and review workflows instead of more terminal tabs.

github.com · 12 min · Agent Orchestration · Agent Workflows · Coding Agents
11-10

Super-agent Design: Router/Worker Split for Always-On Agents

The DeepSeek Harness team applies classic systems ideas to agent architecture. A long-lived Router session handles all human interaction and task dispatch but never executes work; context-isolated Workers run tasks concurrently and subscribe directly to event Threads, so follow-up events bypass the Router. Environment events (Triggers) are filtered by code fast paths before any model call, the context window is treated as a cache with Inline/Spill plus multi-hop Chain of Recall, and an external Watchdog session checks whether Workers are actually making progress. In a synthetic ablation, the two-level design finished tasks at p50 12.4 / p95 15.7 minutes versus 20.4/29.5 for a single session with a subagent tool; removing pass-through raised Router cache-read tokens 37%. The authors flag synthetic workloads and unresolved cloud/endpoint security.

11-09

Loops and Graphs: stop babysitting agents, approve only the last step

Two layers, split cleanly: a loop lives inside a node (produce, check, correct until green), a graph lives between nodes (what work exists, what runs in parallel, where results return). The argument is that a loop without a check that can actually fail is just a scheduler, and that most teams build the work first and bolt a model-based review onto the end. The graph layer gets four node types — splitter, worker, code node, gate — plus two return edges: a short correction edge that fixes the current run, and a long learning edge that feeds an accepted result back into the splitter's brief as a constraint. Gates should open by blast radius, not model confidence, with irreversible work in a lane that never opens. Written for engineers still approving every step by hand.

11-09

Plan With Code, Not Prose: pstack Pt. 2

Part two of poteto's pstack guide covers the front half of the agent workflow: research, planning, prototyping, and architecture. She argues agents mostly fail for two reasons — under-specified intent and missing context — so she has the agent restate a reported problem in its own words before touching code, then uses /how, /why, /recall, and /teach to rebuild runtime mechanics and historical motivation from git history, PR comments, Linear, Slack, Datadog, and Sentry. Planning happens in code rather than prose: README-driven development, a tutorial written before the API, throwaway prototypes driven by /control-app with screenshots and timing measurements, and /architect, which spins up parallel design candidates across model families and has a different model cross-judge them. Escape hatches like any, forced casts, or the same workaround at unrelated call sites count as proof the design is wrong. Aimed at engineers already shipping agent-written code to production.

06-27

Stop Being the Loop: How to Make Claude Work While You Sleep

Boris Cherny, who built Claude Code at Anthropic, no longer writes prompts by hand—he writes loops. This guide explains what a real loop is: a small system that runs Claude repeatedly until a job is done, complete with self-checking, state persistence, and automatic stopping. Unlike cron jobs, loops contain a decision-maker (Claude) that can adapt mid-stream. The article covers Claude Code's /goal (loop until done) and /loop (repeat on a schedule) commands, and provides a paste-ready charter template with sections for goals, work sources, work instructions, self-verification, memory, and stop conditions. Ideal for engineers transitioning from prompting to building persistent, autonomous AI workflows.