Agent cost benchmark on 114 tasks: money goes to retrieval, not reasoning
Coworker_ai · reddit · 2026-09-04
The Coworker team ran 114 tasks with and without a memory layer in front of Claude, same agent and prompts:
- 66% overall cost reduction; 89% cheaper on Jira/GitHub/Slack lookup tasks
- Surprise: expected wins on hard reasoning tasks, actual wins on retrieval-heavy tasks
- Key insight: agents re-derive yesterday's query every run; nobody splits retrieval spend from reasoning spend, so it just inflates the bill
Takeaway: memory/caching layers pay off primarily by cutting repeated retrieval, not reasoning.
Related event: Memory Layer Cuts Agent Costs 66% Across 114-Task Benchmark(2 posts)→
More from coding & agent
- Coding agents plan well but still struggle with dexterous, contact-rich robot tasks — ManlingLi_ · 2026-10-21
- Omarchy 4.5 RC1 ships with 50 languages, AArch64 support and an agent console — mark_k · 2026-10-21
- Opus 5.5 distills a 30-min pi-durable lecture into a 7-min narrated explainer video — solyarisoftware · 2026-10-21
- OpenSearch MCP Server lets agents query and analyze Wazuh security logs — modelcontextprotocol · 2026-10-21
- chaos-index MCP server wraps StupidAPIs with no API key required — modelcontextprotocol · 2026-10-21
- UC Berkeley open-sources SkyDiscover, an AI framework that evolves better algorithms and builds full systems — CShorten30 · 2026-10-21