Harinder Takhar
6.8K posts

Harinder Takhar
@makingmusic
तत्त्वमसि, realizing it in this life through paytm, paypay and woodworking.

Everybody who has ever built software is experimenting at an unprecedented level and the best ones amongst us are pushing the limits to understand how the world will work tomorrow. This is fun @deedydas thanks.

Karpathy's Autoresearch pushed my vibecoded Rust chess engine AI from "expert" to a top 50 grandmaster, a #311 chess engine. It ran over 70 experiments on its own and tried to hill climb to the top ELO score it could, landing at 2718!



prompt get many PRs to prod autonomously using codex subagents: Run a continuous prod-to-green swarm loop. Keep the immediate blocking task local. Use a small stable set of persistent subagent lanes: 1. prod monitor 2. staging shepherd 3. current/newest pathology investigator 4. current fix worker owning the patch/worktree Manage subagents actively: - Give each agent one durable role, one owner lane, and one concrete output contract. - Reuse agents with send_input when new evidence appears; do not respawn unless the lane is genuinely new or the old agent is stuck. - Treat new information as first-class work: when the main thread or another agent learns something material, decide explicitly which existing agent should receive that delta. - Ask agents to report in a compact stateful format: current belief, what changed, confidence, next action, blocker if any. - Require monitors to stay persistent and report only on meaningful state changes, not one-shot summaries. - Do not close or interrupt agents casually; only do it when the lane is complete, superseded, or clearly mis-scoped. - Prefer fork_context=false for narrow review/monitoring tasks; use fork_context=true only when continuity from prior lane context is actually needed. - Poll sparingly. Wait only when blocked on that agent’s result. For every delegated task, require concrete outputs only: - evidence - likely root cause - smallest failing test - smallest safe fix - focused validation - commit SHA if code changed - residual risk - whether this creates NEW_PATHOLOGY or is same-family noise If NEW_PATHOLOGY appears, keep existing monitor lanes running and spin one fresh investigator + one fresh fix-worker lane for that pathology. Optimize for the fastest safe path to prod green. Keep going until prod to green

your edge is whatever you know that the models don't know

I'm going to rile up the trolls with this right now but I am working on 3 different big projects simultaneously across 15 @conductor_build sessions all the time. In the last 7 days I'm averaging 17k lines of code per day, 35% tests, all thanks to gstack. (All mornings/nights/weekends on top of my real job at YC) I ran /retro (from gstack) on all three projects and this is what came back:

@toddsaunders @conductor_build Using the /playground skill for architecture Spending most of your time in plan mode

TIL about the /btw command in claude code

🚨BREAKING: Scientists tracked a bird that flew 8,425 miles (13,560 km) without stopping even once — the longest non-stop flight ever recorded.




