Skip navigation
It is becoming f*cking dangerous: Microsoft Research made Claude Code 23.6% cheaper by letting it…
0:00 / 0:53

It is becoming f*cking dangerous: Microsoft Research made Claude Code 23.6% cheaper by letting it…

plusxbt

13K followers

Visit on X
306 likesDownloadOriginal post

28,557 views

It is becoming f*cking dangerous: Microsoft Research made Claude Code 23.6% cheaper by letting it learn from its own past sessions, and 6 more repos stop it from repeating your old mistakes 1. Install TokenEconomizer and let it learn your repo from past sessions, a fact loads at session start once 3 sessions confirm it, and contradicted facts get dropped: 2. Run loom, ClickHouse's self-hosted MCP server, so what one agent learns the next one can find, its Claude Code plugin loads context at start and records each session: 3. Run npx @globalmpc/nomore to find the instructions you repeated in 2+ sessions and turn them into CLAUDE.md rules that survive compaction: 4. Add stalemap to catch dependencies that changed after the model's training cutoff, it writes a ranked list of changes into CLAUDE.md before the agent starts: 5. Install landmines to find open GitHub issues filed against the exact versions in your lockfile, the agent's edit stops once and gets the workaround from the thread: 6. Open Rulebook to see every CLAUDE.md and AGENTS.md across your projects, it flags diverged files, broken Carl Noel and files over 200 lines: 7. Give small jobs to lean agents, subagents that start at about 7.4k tokens instead of 131.1k because they skip CLAUDE.md, the skills list and MCP tools: TokenEconomizer on 277 SWE-Bench tasks: 18.4% to 30.9% lower cost across 5 setups, cheaper in all 30 cells All 7 repos are less than 6 weeks old, Rulebook needs macOS 14+ Full guide in the article below

10 Comments

Comments from the original post on X (Twitter)

  • Merr's profile picture

    @MerrCat001

    Well the hard part is keeping it lean, im feel like im forever chasing this.

  • Kirill's profile picture

    @KirSolo193083

    the repo list is useful

  • Tolik's profile picture

    @TolikWeb3ep

    turning session history into hard rules and running lightweight subagents without context bloat turns development into ultra cheap and fast pipeline

  • Kirill ZxcQwe's profile picture

    @Kturt1

    Fresh memory and context pruning tools turn AI agents from forgetful rookies into cost efficient pros

  • The SI Maximalist's profile picture

    @thesimaximalist

    Repo memory needs a way to retire stale facts. After a refactor, yesterday’s helpful note can become tomorrow’s repeated mistake. Linking notes to the code they describe gives you something concrete to check when it changes

  • Kagi's profile picture

    @Kaggii_

    all 7 repos being under 6 weeks old is worth a read before installing. several of them write into CLAUDE.md, so check what they change first

  • Novlyx's profile picture

    @Novlyx

    Learning from past sessions only pays off if the bad lessons get cut too. Otherwise it just gets cheaper at repeating the wrong thing

  • Luxury Cache's profile picture

    @luxurycachecom

    Caching?

  • AIfa · CODE Eternal's profile picture

    @CODE_AIfa

    Learning from past sessions is great until one session was poisoned — then the agent keeps "remembering" the attacker. Worth pairing with a guard at write time: PADAM NOCTURNE holds planted instructions for review, masks keys before storage, and lets you forget a single record provably.

  • Frankie Wu's profile picture

    @frankiewuxyz

    the cli is quietly becoming the agent's memory. CLAUDE.md, session logs, skill files. i run most of my week through it. the biggest gain wasn't any single tool, it was the agent remembering what i corrected last time.