Glean 拾遗
Recent picks

4picks · 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.

11-13

Two Agents, One Record: Git-Style Branches Instead of Row Locks

Two agents doing read-modify-write on shared memory lose updates silently: because an agent writes back the whole record it read, a stale save can reset a field another agent just changed, and both report success. The post walks through the three usual fixes—single writer, row locks or optimistic version checks, last-write-wins—and argues all fall short for multi-step tasks spanning several records, since none produces a reviewable diff or an audit trail. It then demos Omnigraph, an MIT-licensed graph database in Rust with a Git-style model: each agent commits to its own branch, merges compare against the merge base, and a stale branch is rejected with divergent_update instead of publishing. The walkthrough uses Claude Code and Codex, covering @key stable IDs, --snapshot reads of old commits, and a Cedar policy with per-agent tokens returning HTTP 403 for writes to main. Useful for anyone building multi-agent shared state, though the headline's 1,000-agent scale is never shown—the demo is one developer and two agents.

x.com · 11 min · Agent Infrastructure · Agent-Memory · Multi-Agent
11-12

1,000 agents, no researcher: a self-organized Lean proof of Fermat

A thousand agents running an early build of Claude Fable 5.2 formalized Fermat's Last Theorem in Lean in 28.7 hours, without the Prove2Me coordination platform or a researcher steering the work. Each agent got its own network-isolated sandbox, a Git branch, a shell, Lean plus Mathlib, and access to a shared Postgres database. The lead model spent its first ten minutes creating five tables to announce orders, assign statements, collect notes, queue subtasks and log finished work; helpers then claimed tasks and reported progress through them. After most agents judged 47 of 60 statements infeasible, the lead reallocated nearly everyone to the remaining ten, and later built a 2,418-node task tree up to ten levels deep for Langlands–Tunnell. The post details the safety stack—sandboxing, a blocking classifier on every tool call, and offline monitoring—and concedes that agents still misjudge difficulty and quit early.

claude.dev · 23 min · Agent Harness · Ai Safety · Formal Verification
11-10

Choosing a Claude model: dial effort before you switch tiers

Anthropic's own guide to picking among Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2. The argument: adjust effort level before switching models, because a higher tier at a lower effort is often cheaper per finished task than a smaller model at max. Evidence is concrete — Opus 5.5 at high beat Fable 5.1 at max on Terminal-Bench 3.0 (58.9% vs 58.0%) using half the tokens; on FrontierCode 2.0 Fable 5.2 at high roughly matched Opus 5.5 at max (45.3% vs 45.5%) at about half the cost per task. On a SWE-bench Pro subset, running everything at low and re-running only failures at high reached roughly 97% for $0.17 per task, against 95.3% for $0.29 with everything at high. Also covers cache-read economics, preserved-thinking rules when swapping models, subagent YAML, and migration gotchas that return 400s. For engineers building agents or budgeting token spend.

claude.dev · 20 min · Claude · Cost Optimization · LLM