Building with coding agents. Writing about AI engineering, verification, agent-native codebases, and what survives production. DeFi since 2019.
East Coast, US
Joined January 2015
- Tweets9.4K
- Following808
- Followers2.7K
- Likes10.4K
Pinned Tweet
CPVO: Cost Per Verified Outcome for Engineering Organizations
The metric that answers a simple question: Is your AI investment actually paying off?
A deep dive into the methodology and how to measure it ⤵️
sawinyh.com/blog/measuring-a…
Reset git history between coding-agent eval trials, not just the working files. Anthropic observed agents gaining an unfair advantage from previous trials' commits. A clean checkout can still contain the answer in its history.
Giving a coding agent file paths instead of full files keeps context lean, but adds retrieval work at runtime. Preload the essentials; let it explore the rest. Context size and time to an answer are different optimization targets.
An agent eval that only checks whether search happens when needed can reward searching on every query. Include cases where the agent should answer without a search, or your score can improve while unnecessary tool calls multiply.
MCP agents can filter tool results in code before the model sees them. But running that code needs sandboxing, resource limits and monitoring. Evaluate the token savings against the runtime you now have to secure and operate.
Nick Sawinyh retweeted
Cold email was very effective until around 2020.
When I started in sales in 2016, I would send around 100 lightly personalized cold emails a day and get around a 5% positive response rate. Around 30% of those resulting meetings converted to sales. The math worked.
Then in COVID, everyone who was doing outside sales was forced to pivot to inside sales and used automation platforms like SalesLoft and Outreach as a spam cannon.
This was also the ZIRP period when SaaS companies grossly overhired on their sales teams and thus with more sellers than leads to go around, sellers resorted to spray and pray email tactics
The final nail in the coffin for cold email was agents that can send emails. If you receive a cold email today, there is probably only a 10% chance that a human put actual thought into it
Cold outbound as a strategy still very much works. But it needs to be hyper personalized, and is only very effective on platforms like X and LinkedIn where potential buyers can learn more about your personal brand before engaging
Smart Money bought a token. Is anyone buying in behind them?
Zatto counts the new buyers in the 10 minutes after each Smart Money buy, then tells you what buys at that level did 24h later.
Built on the @nansen_ai API
zatto.nsawinyh.workers.dev
github.com/sneg55/zatto
What it has measured so far, 542 Smart Money buys on Base:
under 5x the token's normal buyer rate, copying is a coin flip
5x and up, 20 of 25 were higher at 24h, median +29.6%
One row per token minute, so a cluster of wallets buying together counts once, not seven times.
A run goes every 8 hours on cron. If you want one now, a scan costs 1 USDC over x402, signed in your own wallet. The app never holds a key.
Method, constants and the measured numbers are all in the README.
github.com/sneg55/zatto
Nick Sawinyh retweeted
anthropic released a new prompt that eliminates claudese:
platform.claude.com/docs/en/…
I paid 12 x402 sellers for the same thing: the USD price of BTC.
Only three answered.
The rest took the money anyway. One gave back HTTP 500. Another sent real BTC data as a JSON array, breaking the object contract it publishes itself. A third answered a question I hadn't asked.
You can read every listing price for free. You cannot read whether a seller delivers, and a spend cap only ever sees the price.
Kenyaku pays once to find out, then skips those nine forever. Cost per useful answer went from $0.008333 to $0.001000, and the record of who failed lives in @sibylcap memory.
Settled $0.025 on @base, reconciled against the USDC Transfer log. Code and evidence:
github.com/sneg55/kenyaku
Kenyaku, in three minutes.
Session 1 starts with an empty store. It pays all 12 sellers and finds 9 of them broken.
Session 2 is a fresh process. It reads the spend policy out of @sibylcap memory and pays exactly one.
Then I delete the store and run the identical command. Back to paying 12.
youtu.be/HYzSgal7EdU
Nick Sawinyh retweeted
How to run founder sales without destroying your build time:
1/ Calendar lockdown. Batch all external calls into Tuesday and Thursday afternoons. 1pm to 5pm only. Monday, Wednesday, Friday are zero-external-call days for product focus.
2/ Qualify brutally on the intake form. If they don't have the budget or a live problem your v1 actually solves, drop the call. Your hours are more expensive than equity right now.
3/ Sell the raw product yourself. You built it, so you know why the competitor's architecture breaks. Ask the uncomfortable questions a junior rep is too scared to ask.
4/ Document every objection after the call into a single Google Doc. By month 3-4, you have the exact playbook ready for your first AE hire instead of training them off vibes.
5/ Record all sales calls, wins and losses are both equally as valuable. Use Granola to add the notes into Hubspot.
6/ Let it be known company-wide that sales is the company’s 1st priority. This will signal to all that you’re in growth mode and unite under 1 goal - revenue. Be clear about this goal and give 1 quarter a full sprint in theme.
Founders who try to be polite to every tire kicker end up with a full calendar and zero revenue.
Nick Sawinyh retweeted
GTFS packages a transit schedule as plain-text tables linking agencies, routes, trips, stops, and times. GTFS-Realtime adds delays, vehicle positions, and alerts on top.
its-feed.com/articles/what-i…
#GTFS #TransitData
Built VerdictLock for Telegraph Track 2, a URL_SCAN scoring module.
On the 40-case benchmark the Canonical Script gets 20/40 ordering wins (margin 0.0133, spread 0.1354). VerdictLock gets 40/40 at 0.9169.
Repo + harness: github.com/sneg55/verdictloc…
@Telegraphprotoc
I built an agent fleet over 41 federal data feeds and kept the model out of every decision that matters. Six deterministic rules hold the gate; Gemini never touches a verdict.
Why, and the feed that started it:
#AllThingsAgentic
sawinyh.com/blog/conformance…
