Five agents each re-reading a 4,000-token brief is 20,000 tokens before anyone does any work. One agent reads it once. Default to fewer, fatter agents until you can show the number that says the split pays. New edition below!

Just spotted Byte-Sized Design at #6 on Substack's Rising in Technology chart. 🎉 Genuinely floored, this newsletter is nothing without the 43k+ engineers who read, share, and reply every week. Thank you for trusting us with your inbox. More deep dives and roundups coming. Onward.

Stripe replaced Envoy with their own Go proxy and halved both CPU and latency. At their scale that's ~3,000 days of inter-service latency saved every day. In this edition we explore why they left Envoy after 10 years, and the connection-count trap that pushed them:

It takes four vendors to summarize a phone call. One transcribes. A regex redacts. An LLM summarizes. Compliance reviews all three. Three retries per layer, three layers deep: 3³ = 27 requests from one failure. New edition, together with @AssemblyAI

GitHub's remediation list from Monday's outage has five items. Four are bug fixes. The fifth is "review retry limits and backoff behavior across gateways and clients." That's the only structural one, and scoping it as a review undersells it. Retry policy is a capacity decision that nobody owns. One team adds a retry to a flaky call. A client team adds retries because laptop networks are bad. Both correct in isolation. The sum is a load multiplier that only activates under stress. The autoscaling misconfiguration gets fixed in an afternoon. The retry architecture is still there waiting for the next trigger.