Addy Osmani’s ‘Loop Engineering’ this week captured what Boris Cherny, who built Claude Code at Anthropic, said plainly: “I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”

Half the developer timeline nodded. The other half asked what it actually means in practice. Both halves were software engineers.

We’re not software engineers. We run analytical intelligence at Manthan Intelligence. For the past three months, we’ve been building exactly this — 46 agents, 23 agents per human, running on scheduled loops across a knowledge graph of 102,000 entities covering companies, investors, and relationships.

The failure modes were identical to what the discourse describes. The fixes were slightly different.

1. The loop without feedback is a machine for generating confident mistakes.

Our NFX classification pipeline ran thousands of companies through three separate loops before we stopped. Each run failed a pre-committed 0.96 precision gate — not randomly, but with a different error class each time. The third run’s errors were entirely disjoint from the second’s. That’s not a tuning problem. That’s a method ceiling.

Pre-committed gates before run 1 are the only thing that kept wrong labels from silently entering a knowledge graph that 46 downstream agents read.

The rule we locked in: set the gate before you run. If you negotiate the stop condition after seeing the results, you’ve already lost.

2. The expensive resource shifted — but it’s not just money.

Uber capped Claude Code at $1,500 per person per month after burning their entire 2026 AI budget in four months. We took a different path: zero external API spend by default. Every scheduled loop runs on our Max subscription. No run authorises spend without an explicit flag and a cap.

The cost discipline isn’t frugality. It’s about keeping loops interruptible. A loop you can’t afford to stop is a loop you’ve already lost control of.

3. The anchor file is the loop’s memory.

The Tattooed Ralph pattern made this explicit: reset context to a fixed anchor file each iteration, rather than letting the conversation grow. We’ve been running this pattern for three months without calling it that. Our system context file is the anchor — calibration notes, method boundaries, failure modes from previous runs.

The loop doesn’t accumulate context rot. It accumulates calibration.

4. Model identity is self-reported, not attested.

This one isn’t in the discourse, and it burned us.

We designed a cross-model A/B test and reached for a Max-alias subagent as the cheaper arm. Caught it before running: the subagent’s model identity is self-reported, not verified by the infrastructure. Production prompt-chain parity fails by construction. Running it would have generated 18 assessments we believed were from Model B but couldn’t verify.

Loops give you scale. They don’t give you experimental validity for free.

5. The skills are the asset. The loop is plumbing.

The clearest articulation of this is visible in the skills ecosystem: a loop that calls sharp, tested, named skills compounds. A loop that re-derives its context from scratch on every run just burns money. Matt Van Horn’s last30days skill is a worked example of the first approach.

Based on Manthan’s internal calibration data, we have 12 named analytical lenses — each calibrated against 1,100 scored historical outcomes, with a weighted accuracy at 65.1% and improving weekly. The loop (cron-based, durable, git-backed) runs them. The accuracy comes from the skills, not the scheduling.

The discourse this week is about coding agents. The same architecture works for analytical intelligence. The same failure modes appear. The same fixes apply.

The interesting thing isn’t that loops work. It’s that the feedback inside them is where all the difficulty lives — and where the compounding begins.

— Manthan Intelligence

This analysis draws on Addy Osmani’s Loop Engineering, Uber’s cap on Claude Code spending, and the Tattooed Ralph agent pattern. Human editorial oversight applied.

This analysis is informational and does not constitute investment advice, a research report, or a recommendation to buy, sell, or hold any security.

Charaka Notes by Manthan Intelligence. Subscribe