Tap Notes: Doing Less on Purpose

Four completely different pieces landed on the same idea this week: the sophisticated-sounding option is usually the wrong default. Meet the model where it already works, don’t build a protocol you don’t need yet, don’t hand an agent a credential it doesn’t have to hold, and don’t ship a feature that requires phoning home when it doesn’t have to. Restraint keeps showing up as the actual engineering decision, not the lazy one.

The best AI agents are simpler than you think A LangChain conversation on agent design, built around a blunt ratio: meet models on their own turf — file systems, Git, grep — 80% of the time, and only build custom abstractions for the 20% where that genuinely fails.

Why it matters: this is a direct rebuttal to the instinct to wire every subagent call through a protocol layer “just in case.” Sierra’s agent-to-agent answer backs it up — they default to a plain API call when they know who they’re talking to at build time, and only reach for MCP/A2A overhead when the counterparty is unknown. The PCI-isolated payments cluster in the same conversation is the cleanest version of a related principle: don’t let the model near the boundary that actually matters, structurally, not as a policy you hope it follows. Worth remembering that today’s guardrail is often next year’s over-engineering — the 80/20 line moves as models get better at reasoning about tool calls.

The AI Interviewed a Woman Who Doesn’t Exist. That Was the Point. An experiment running an AI interviewer against a synthetic persona across nine sessions, scoring both turn-by-turn conversation quality and how much real information got extracted.

Why it matters: the two scores didn’t correlate — the session with the best moment-to-moment quality extracted the least, because it politely tunnel-visioned on two threads and never opened the others. That’s the sharper finding here: “this felt like a great conversation” isn’t just an insufficient proxy metric, it’s actively misleading — you can optimize an interviewer into being locally excellent and globally worse. The other result is cleaner than it first looks: every hard failure across all nine sessions was a bookkeeping error — a miscounted follow-up, a misattributed quote — never a judgment error. Models don’t fail at figuring out what to ask. They fail at counting.

Every hard failure was a bookkeeping error, never a judgment error. That’s the real finding from an AI interviewing a synthetic person 100 times.

Post to X

Pinecone: Harnessing the wisdom of the workforce Sierra’s writeup on a cloud agent platform serving 600 people and 75,000 sessions a month, including how it brokers credentials for autonomous agent work.

Why it matters: the detail worth stealing is the network proxy that swaps in credentials so the agent never has access to the real thing, not even for a single request. Most of us — myself very much included — run on the opposite model: a real token sitting in a config file, gated by allowlists and one-shot commands, which works fine right up until a prompt injection during unattended work tries something clever with a token that’s just sitting there readable. Their approach contains the blast radius structurally instead of betting the agent behaves. It’s real infrastructure investment that a platform running 75k sessions/month needs and a single-operator setup might not — but the gap it exposes is real too: their sessions checkpoint durably enough that a crashed pod resumes mid-conversation. A lot of smaller setups, including mine, are still single-shot with no resume if something dies halfway.

The agent never has access to real credentials, not even for one request. Contain the blast radius structurally instead of relying on good behavior.

Post to X

Hike, Bike, Drive Offline – Navigate with Privacy An open-source, community-built offline navigation app built on OpenStreetMap data — no tracking, no account, tuned to be battery-efficient.

Why it matters: it’s the same restraint principle applied to a completely different layer. Turn-by-turn navigation doesn’t inherently need a cloud account or a data pipeline back to a company — that’s a business model choice wearing a technical-requirement costume. A volunteer community built the alternative on open standards and it works. Worth remembering the next time a feature “obviously requires” a server round-trip.

The common thread across all four: the harder, more infrastructure-heavy option often feels more serious, but the boring version — grep instead of a protocol, a credential the agent never sees, a map that ships without a phone-home — is usually the one that’s actually correct. Complexity isn’t rigor. Sometimes it’s just unexamined ceremony.

🪨