Filesystem based context engineering TLDR; we gave our agent a computer and loaded it with files about the account. It decides what it needs to read dynamically. This brought our token cost down from ~$1.00 per call to ~$0.25 (Claude Opus 4.5). The output seems better too (based on vibes, evals on the way). LLMs are great at the command line, why not let them use it? Example: [Sandbox Context] Found 20 Gong call IDs for account [Sandbox Context] Generated 20 Gong call files files written to sandbox . └── gong-calls | └── 2025-11-28-internal-sync.md | └── 2025-11-26-acme-vercel.md | └── 2025-10-24-acme-vercel-session-2.md | └── 2025-08-19-acme-vercel-sysops.md | └── 2025-05-06-acme-v0.md | └── ... (15 more) └── slack | └── internal-channel.md | └── external-channel.md └── salesforce | └── opportunities | | └── current.md | | └── 2025-10-31-acme-aws-new-business.md | | └── 2025-09-22-acme-v0-10-seats.md | └── account.md executing command { "command": "cat", "args": ["salesforce/opportunities/current.md"] } executing command { "command": "cat", "args": ["slack/external-channel.md"] } executing command { "command": "grep", "args": ["-i", "visit\\|signing", "gong-calls/2025-11-28-internal-sync.md"] } Blog post coming soon!