Drake Evans (@drakeevans) profile picture

founder @withAUSD. economics & code. formerly: fraxlend author @fraxfinance, engineer at @ADP

New York, NY
Joined January 2014
Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)Photo by Drake Evans (@drakeevans)
Everyone QT'ing OP but this reply is the real way things will go. 3 Predictions: GUIs are here to stay but become custom AI harnesses over time APIs move to building blocks over pure orchestration. They are more important than ever because they are what underpins the MCP/CLI tools etc. Exportable building blocks will become paramount. They let people build their own UIs/orchestration in whatever way they want. The UI is just an opinionated configuration of core capabilities with a display/discovery/harness layer on top. More and more people will want to build these pieces into their own UI or more increasingly their own harness All of this is why you must own the metal to own your own destiny
Replying to @SamBroner
FWIW, my preference would be that every SaaS offers three great experiences with a shared data and permission model: 1. Great app 2. Great APIs/MCPs so I can bring my own agent 3. Great deep linking & exportable components — e.g. help my agent show me “just one thing”
1
1
14
2,065
Democrats delivering CLARITY was not on my bingo card
🚨SCOOP: House Democrats are discussing plans to expand the Crypto Investor and Consumer Bill of Rights, @RepMaxineWaters’ amendment to the Clarity Act in 2025, into a broader crypto “market structure” bill next Congress. The proposal would fold in Democratic Clarity Act provisions originally offered by Waters in the House (which were not adopted), as well as provisions Senate Democrats were able to agree to with Republicans before the September cloture vote, according to three industry sources who have spoken with Democratic staffers. Some of the proposals under discussion are said to include giving the @SECGov more of a role in supervising the crypto markets, and a more limited role for the @CFTC. Additionally, it’s likely that further ethics provisions will be offered to restrict the crypto activities of President Trump and other federal officials. Republicans are calling this a “messaging bill.”
1
1
9
395
If your company thinks the 512gb m5 Studio is overpriced you're ngmi
3
263
I havent yearned for the mines (coding by hand) for years but the final vision of solar makes me miss those days. LSP near and dear to my heart because I started building one as the options were so bad back in 2020, which then led me to static analysis tools AST, CSTs etc etc. So much great learnings came from those efforts
Replying to @gakonst
Solar has an LSP! Check it out here. It's really epic.
4
255
Rex + cua spaces would be so nice, every session a sandbox thats either a container,vm,or full machine. Native apps + TUIs all within that nice window manager. Teleport between sessions. Wen
4
303
Drake Evans retweeted
Agora just got its OCC charter approved. That’s a major step for regulated, on-chain stablecoins. @drakeevans @withAUSD @therollupco
1
7
432
Serif fonts are so back!
Introducing Geist Serif.
4
590
The bigger problem we will have to confront is that serving your product to users is the exact thing that allows them to build it themselves (similar to distillation)
This is the most savage thing I could imagine. A competition to Reverse Engineer your competitor as open source. The internet is getting wild...
1
3
310
My agent just reads the skill file then does it anyway lmao
The skills spec is so broken
1
183
Drake Evans retweeted
forget ARR, this is it
Replying to @trycua
@trycua is like lil wayne in 2007-08, in the middle of generational run and you know its happening
1
6
383
Wait until you hear about harness outside sandbox, full session durability (w/ forking), and boxes that are "ephemeral" but you can come back to forever
My personal sense is: Cloud Agents ("ephemeral", per-PR cloud container) < Local dev < Cloud Agents ("devbox", long-lived, laptop-isomorphic) Unfortunately - many services err towards selling you the former (Cursor, Claude, Codex, Conductor). For the "devbox" style Cloud Agents - I'd recommend Claude remote-control or Paseo. I use the latter. (Cursor-over-SSH or Codex-over-SSH work too ofc) Why is ephemeral so much worse than devbox? There's just a lot of types of debugging, deployment, testing workflows that are SWE, but not exactly "writing a PR" that go much worse w/ ephemerality. OTOH - for devbox-style, the answer is always "just mimic what your laptop does". Some examples: - I want to do `terraform plan` and `terraform apply` with my agent. This means I need to give my agent AWS perms. - How do I do this with a ephemeral cloud agent? Err... I mint long-lived access keys for it - which my infra team doesn't support. Maybe I could copy my personal creds over? What do I do when I have multiple AWS accounts? - How do I do it locally? Easy, `aws sso login`. How do I do it on a devbox? Easy, `aws sso login`. Much like my laptop, I might log into multiple AWS accounts. - I want to iterate on an app that has a DB. How does this work? - Ephemeral: Maybe re-seed the DB each time? Maybe use a novel DB provider w/ forking - but no I need a real DB like Aurora Serverless. I'd rather do PG, so I guess I build some sort of seeding. - Devbox: It's just like using a laptop - long-lived DB, gets interesting state over time. [Related: a lot of the `npm run dev` local-dev stuff is very nice, doing quick preview deploys instead of exposing localhost is bad - it's just never as fast as just HMR. Devbox + tunnel for localhost + agent docs on how to use the tunnel is really sweet] A few others I won't detail: - I want my agent to read my datadog logs to help me debug an issue, how does that work? - I want my agent to set up a new Cloudflare subdomain / worker - I want my agent to be able to hit internal services on my Tailnet - I want an agent-only journal & task-list that is on a separate cadence from my repo (ie: often failed PRs that do not merge still want to write to the journal, and this would make too much PR/commit noise) - I want to build & smoke-test & deploy docker containers, and ephemeral agents run into dind issues. - I want my agent to help me zip up some files, inspect them for a while + put them somewhere else. (likely AWS access bottlenecked?) In general, the core problem ephemeral cloud agents run into is SWE involves a lot of tangential poking through platforms + long-lived access. You want a permissions model that mirrors a laptop (ie: all developer-relevant perms), not a Github CI runner (ie: only app-deployment relevant perms). [A secondary problem is: local state is more nice than it is harmful. It can be harmful for reproducibility, but far more often it's nice for journals, dev DBs, static files] [Another problem is: there's also security concerns about using 3P sandboxes, your infra team might drag their feet, and long-lived boxes are super trivial to self-host - just make an EC2 VM] If you have a phenomenal devprod team, one answer is: "they've provisioned a lot of short-lived workload identities, that you can assign to ephemeral containers". But also ugh - different engineers on different teams have different access - and might not even want all their ephemeral boxes to have complete access? How do you let people configure this smoothly? With devboxes it's easy - folks often just have N devboxes, with different configurations.
1
6
393
As the daily requirements for physical labor decreased over previous decades dedicated time for physical exercise became more and more important. We are now entering the same phase with intelligence. It will become extremely easy for our brains to live a sedentary life and atrophy
3
1
24
1,383
Working with @HadickM and Dragonfly team has been one of the best decisions we made at Agora. Well deserved and by far one of the best VCs out there.
Over the last four years during my time at Dragonfly, we’ve consistently discussed who we look up to, how we want the firm to evolve and exist many years from now, and what is the legacy we want to leave behind. The Benchmark partnership model and their ability to see around corners is one we have consistently come back to as a North Star, and I’m excited and proud to continue to help lead Dragonfly into that future, alongside Haseeb, Bo and Tom, as we grow and continue to back the best founders innovating across money, markets and intelligence.
1
20
2,110
1/ The @SECGov rightly recognizes that a token can be a non-security asset. The next step is a clear rule for when a token can trade freely. @ThaniaCh explains why it matters for exchanges and markets. Watch Below ⤵️
1
7
23
1,446
It’s pretty obvious at this point that the only reason that model companies give us access to anything is to subsidize their compute which they use for inference / eating whole industries / solving math/cancer etc. we will never again have access to the frontier
2
5
520
Drake Evans retweeted
AI agents can’t buy a license or open a bank account. They will route through whoever has the best on-chain economics. @drakeevans @withAUSD @therollupco
2
1
3
348
Drake Evans retweeted
Banks have a mandate to get to 24/7 markets. Stablecoins make that possible without waiting on traditional rails. @drakeevans @withAUSD @therollupco
2
1
9
720
Drake Evans retweeted
Drake Evans on the future of stablecoins and AI agents. Efficiency, 24/7 markets, and real utility will decide who wins. @drakeevans @withAUSD @therollupco
2
1
9
374
Insane! World will transform in the next 12 months: 1 yr ago we were using opus 4.1 and GPT 5
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
3
259