Shared Memory for 3 Parallel Research Agents Saved Nothing: A $1.11 Post-Mortem
quiet_flare · reddit · 2026-10-20
The author tested whether shared memory (page digests with URL + content hash, plus shared prompt prefix for cache hits) could cut costs when three research agents run in parallel. Result: Arm A (no sharing) cost $1.11 with 16 fetches; Arm B (shared memory + prefix) cost $1.09 with 13 fetches — no savings. Logs show why: agents started 6ms apart so recalls always hit empty memory; simultaneous first requests all missed the prefix cache (and the harness put per-agent working directories in the system prompt, making prefixes differ); digests themselves cost 2.8x Arm A's output tokens. Fixes like running one reader before fan-out or staggering starts are runtime-agnostic. Code and data open-sourced at GitHub QuietFlare/mainsheet/bench.
More from coding & agent
- LinkedIn Desk V2 Brings 8 Skills to Grok Bot for Taming Your LinkedIn Inbox — SEO · 2026-10-21
- LF Research: GenAI halves open source coding time, but 25% of PRs are AI slop — nordicinst · 2026-10-21
- How Anthropic reviews Claude Code changes: machines do the reading, humans handle risk tiers — rseroter · 2026-10-21
- Karotte deep dive: an RL environment framework that prevents whole classes of reward hacks — idanbeck · 2026-10-21
- Via AI exits stealth claiming SWE-Bench Verified lead with LLM-optimized compression at half the cost — ashajadeja · 2026-10-21
- Reward hacks often live in the environment, not the weights — idanbeck · 2026-10-21