Why did my agents keep spawning sessions and burning money?
The book tells this one as an incident, then as a fix. Here are the passages, quoted exactly as written.
Telegram is a human channel — when Agent A sends Agent B a message via Telegram, Agent B’s gateway receives it and a new agent run starts. If Agent A then responds on Telegram, Agent B’s new run receives it and starts yet another run. This is the loop.
Read that twice, because it’s the whole failure class. A human channel treats every inbound message as a reason to wake up. Two agents on a human channel are two motion-sensor lights pointed at each other. Each reply spawns a fresh session — a fresh context, with no memory that it is the forty-third iteration of anything — which reads the message, finds it warrants a response, and responds.
On May 16, Echo and HKael got into exactly this loop and exchanged the word “dark” more than fifty times before I intervened. Fifty-plus sessions, each one spun up, briefed, billed, and burned to say one word to a peer that would spin up another session to say it back.
[…]
The fix is the part worth studying, because it was not a patch. Nobody rate-limited Telegram or added a “loop detection heuristic.” Echo wrote a protocol that reclassified the channel itself: bus-first, Telegram-for-humans-only. Peer coordination moved to the
peer_messagesandpeer_inboxtables inagency.db— bus messages don’t trigger new agent runs, so the loop is structurally impossible there.
— from Chapter 11, “Corruption, Cascades, and the Substrate Lie”
Rate limits are not budget caps — one bounds velocity, the other bounds total spend, and an agent rate-limited to ten requests a minute can still burn $500 over a patient afternoon; you need both.
— from Chapter 8, “The Autonomy Dial”
That was a few paragraphs of a fifteen-chapter book.
Buy the full book — $12 →All 15 chapters, EPUB + web, every revision through v1.0 · 14-day no-questions refund
or read the free sample first — the Introduction and Chapter 1 →
More free excerpts:
→ What is an agent factory, really?
→ What should an agent do when a human doesn’t answer?
→ How do you know an agent’s “done” is true?
How this book was made: it was drafted by an AI (Anthropic's Claude) from the primary documents the fleet produced and preserved, under John Whitman's direction. A revised edition (v1.0) is correcting the claims an editorial audit could not trace to those documents.