shae wang

ds @ripple, prev ml @uber
building in crypto: agentic finance+analytics

opinions are my own

[shae wang @shae0x]

Last week, a Head of AI Transformation at a large enterprise told me:

"We're holding off on buying observability/performance eval tooling. Models are improving so fast. Eventually they'll catch their own errors."

That mindset is incredibly dangerous.

Here's why 👇

[shae wang @shae0x]

@dreamsandcode @simonw i've mostly landed on max only when i can't check the output myself, because that's when the extra thinking actually buys me something. for anything i can eyeball in 10 seconds the cheaper tier plus my own review beats paying for reasoning i'm going to redo anyway

[shae wang @shae0x]

use regret minimization to pick your lane
size the investment by how much unfair advantage your have
cap the downside

= compounding growth and fun

[Linda Xie @lindaxie]

Biggest thing my parents drilled into me growing up that I had to unlearn was that it's not okay to take career risks. If there's something I'm deeply convicted in, I go for it. Even if it fails, I learn so much in the process and don't live with regrets on what could have been

[shae wang @shae0x]

this MCP only gives you execution, complete agentic trading needs a layered intelligence stack

assume this simple flow: pull the indicator -> plot its historical distribution -> define bands -> validate

1. foundation: data
- @coinglass_com : rich indicator bands. cross venue metrics incl. funding rate, long short ratio, whale positions (https://t.co/4PlVnHyQix) $29/mo, data dated to 2019
- @Dune : raw cycle metrics only (https://t.co/GL7vdOmlGR) free up to 2.5k credits (~$100)

2. processing
- @duckdb : store data, ASOF joins, sql (https://t.co/iU2999lEgI) free
- agiprolabs' ohlcv-processing: resamples, handles gaps and normalization (https://t.co/sNN0m83GH4) free

3. strategy
- tradermonty's trade-hypothesis-ideator: takes your opinion and generates falsifiable specs, exports strategy.yaml (https://t.co/xtP7IxUxpl) free
- agiprolabs' mean-reversion: band-setting math (linked in 2, free)

4. validation
- agiprolabs' walk-forward-validation: time-series, validation and overfit detection (linked in 2, free)
- tradermonty's residual-edge-analyzer: models what fraction of the return is residual edge vs. baseline exposure, runs decomposition (linked in 3, free)

this gives the mechanics to run a lot of tests, but might have the multiple comparison problem and still catch noise if you run it on all the indicators

@emilylai 's crypto skill might help with theorizing which indicators should matter and scope it down on 1, but i haven't validated this

save and give this list to claude for a one-time install, or please share your own setup so everyone can learn!

[Brian Armstrong @brian_armstrong]

JUST IN: your AI agent can now trade stocks on Coinbase (in addition to crypto, derivatives, etc).

AiFi (agentic finance) is here.

Get started with your favorite harness/agent: https://t.co/O2Od95xdQB

[shae wang @shae0x]

it depends on the use case, the same way human identity does!

think about a store checking your age and taking your card, they’re verifying two unrelated claims from two unrelated issuers. the state says you're over 21. the bank says this card is good for this amount. neither one is you the person, and the store never knows your identity or your balance

agent identity imo works the same way. who built it, who it's acting for, what data it can reach, what tools it can call and when, its model and parameters, its subagent graph, its track record, its token consumption, whether it's been revoked, and who's liable when it's wrong.

no single protocol covers that because no single party can sign it. the model provider can attest to the model. only the deploying org can attest to the mandate. track record only exists after the fact.

and they move at different speeds. model version is static. google's agent identity certs expire every 24 hours. permissions change per session. revocation has to be instant.

i think this sorts itself out as the stakes go up

[shae wang @shae0x]

nothing beats learning these through building by hand and discovering your own mistakes

[Andrew Ng @AndrewYNg]

How have software engineering fundamentals changed with agentic coding? Here is our AI Engineering Skills map for software engineering fundamentals. https://t.co/cnRLj43DLs

[shae wang @shae0x]

their employees are already vibe coding internally

[Alex Turnbull @iamAlexTurnbull]

There’s no fucking chance you’re vibe coding mid market and enterprise software. No chance

[shae wang @shae0x]

@suhburo dope!! any plans to build it for other harnesses other than codex? or across multiple?

[shae wang @shae0x]

why is USDT always below $1 and how are there still billions of transfer volume daily?

[shae wang @shae0x]

@cryptoreine i see, that's hard for mid-high single digits, but excited to see what you bring on!

[shae wang @shae0x]

@adhvfed tbh i couldn't really tell what changed but it does feel easier to follow! but the change doesn't seem to propagate on page flip 🤔

[shae wang @shae0x]

oh this is fun and very humbling. small suggestion for the UI - feels like my eyes have to jump around a lot due to the way the text box moves, possible to make it look like it's scrolling while the text stays in place? if you did it intentionally to mimic real reading, perhaps can blur the previous line so there's a little bit of buffer?

[shae wang @shae0x]

@philhie the system CAN work if you take the time to give it feedback, maintain it, give it context constantly, and govern it well. every one operates differently so you also have to wire up the pieces yourself. it gets rewarding after 1-2 months, which no founder has the time for

[shae wang @shae0x]

one setup and balance for every agent tool is the right product experience

but the framing of "killing" other search tools feels backwards
monid should route across all the search tools based on freshness, index coverage, latency, cost, etc. because search is substitutble, like models

there's a lot of substituble tool spaces to build runtime routing in

[Shengkun Ye @shengkunye]

We just killed Exa, Tavily, SerpAPI, and Brave.

Your agent can now search & fetch any webpage for 100% FREE.

Them: $7 per 1,000 searches.
Us: $0. No subscriptions, no quotas.

Humans search Google for free. Agents shouldn't have to pay either.

Made possible by @Tiny_Fish and @MonidHQ.

[shae wang @shae0x]

@JSON_JEFF @PatchOpsAi that's dope! but that's really awesome that you kept a system of record whether deliberately or not, so much of the data most people can now harvest is either inaccessible to them or gone

[shae wang @shae0x]

@gabrielgtapps @petergyang Woah this is pretty cool, I tried super long-running analytics agents before and had a problem with goal drift and just overall the model burning too many tokens going down the wrong path, what are you running the agents for and how do you solve for it?

[shae wang @shae0x]

@drexthealpha glad it resonates! yeah sometimes just 1-4, the key is to make it lightweight enough so I actually do it, and there’s almost always follow-up questions from me as I discover where I was wrong

[shae wang @shae0x]

i love using voice mode to make claude quiz me after a research session

reason i do this: reading ≠ understanding. LLMs are trained to sound good to humans, but all the work that normally makes something stick happens before sounding good, and if you end up recalling nothing it's just a waste of time

so i ask it to test me instead using voice, because it forces me to recall from my own memory

5 questions i use:
- what were the foundational concepts in <topic>? one or two sentences each
- what did you think before this session, and what changed?
- convince a skeptic that this works (or that it doesn't)
- where in your own work does this apply, and what would you do differently now?
- where does this stop being true? give me a case where it's the wrong call

i run it as a skill when i close a research thread, and either before i pick it up again or at the start of the next session