Founder @ https://xitter.cf/t.co/BGZoFhigki 12+ yrs shipping software. Now building agentic workflows and custom software. Notes from real builds
Hyderabad, India
Joined March 2014
- Tweets2.1K
- Following176
- Followers225
- Likes399
30 vs 46 tool calls is the stat i'd watch over the 4.5s, every extra call is another file read and that gap grows on a real repo
per-token price keeps dropping but agent loops resend the whole context every step, so the bill still climbs unless you cache the prefix
the dedupe part is where it bites, two memories saying opposite things about the same customer and the agent just trusts whichever ranks higher
201s to 31s for the eighth chat is the whole case for sharing the prefix, eight agents on one 28k system prompt shouldn't each prefill it
the prod-only cleanup step is the nasty one, staging never ran it so every later migration got tested against a database that doesn't exist
the $20 tool was never the real cost, the real cost is fixing auth at 2am in an app nobody else uses
the md output is the useful half, paste a substack url into an agent and it often reads nav, paywall text and comments first
price per run matters less than retries, one misclicked cell in a spreadsheet run means paying for the whole run again
banning projects "meant for AI" is the odd part, that rule catches MCP servers and agent tools even when a human wrote every line
the one-in-two-runs break is the scary bit, a single eval pass on the folded version would have looked green and shipped it
human corrections becoming per-client notes is the part that compounds, the n8n syncs just keep the repo from going stale
jina reader and markitdown matter more than the search MCPs here, clean markdown in means way fewer junk tokens in context
the 'waiting on you' hand is the most useful bit, most agent setups bury that state in a log and you notice an hour later
sonnet's has a crowd, combo multiplier and timer while terra's is one kickflip, so the extra $17 bought a game loop, not slow thinking
until that gets better, giving claude code the playwright mcp server gets you real clicks and screenshots it can check its own work against
the $0.003/M cache read is the real story, 2.88M tokens for a dime only works because all 9 reads reuse the same 320K prefix
remotion is the right pick for this since the video is just react, so when arc menu ships a new menu you re-render instead of re-editing
codex looping on checks usually means it has no finish line, a row count and checksum match per table tells it when the migration is done
revoking access is the step nobody tests, pulling the oauth grant doesn't kill the access token the agent already cached until it expires