Mark Li

Building @glasserai to give AI agents an unfair advantage. 2,000+ data tools, 27 providers. Pay per call, no seats.
Before: Lark at TikTok & SureThing Agent

[Mark Li @iammutex]

Your agent is only as good as what you feed it.
Garbage in. Garbage out.
Introducing @Glasserai - it feeds your agent premium data instead.
Ahrefs. Semrush. ZoomInfo. Apollo. PDL. etc.
1,900+ endpoints. One key. Pay per call. No more subscriptions.
Send this one line to your Claude Code, Codex, Grok Bot, or Muse:
set up https://t.co/JRrBd6P3YT

[Mark Li @iammutex]

@daniel_mac8 Half of "using it the best" happens before the prompt. Same model, same prompt, hand it a stale list or a missing file and it looks like a worse model. Most "the model got dumber" takes are really input problems.

[Mark Li @iammutex]

@stretchcloud This is lead scoring in one call: a thousand "is this account a fit?" questions at once. But a calibrated probability on just a company name is still a guess. Give it the company and funding data first. Let your agent pull that through @Glasserai 😎

[Mark Li @iammutex]

@morganjingram > grew their own sdr headcount 106 percent in a year

ai made each touch cheaper, so they bought more touches. cheaper coal never meant less coal burned

[Mark Li @iammutex]

@zeeg a pr count only measures how fast someone can hit merge. the number worth posting is how many of those 7000 got reverted in the next 30 days, and nobody ever posts that one

[Mark Li @iammutex]

@DBredvick the bad data one is the factory i'd run first. scoring qc on top of duplicates and stale enrichment ends up tuning the prompts to fit the bad rows
fix the rows before the prompts, otherwise every 7am pr is grading the wrong thing

[Mark Li @iammutex]

@manojlds Why spend time saving quota? If you're optimizing this hard, maybe the AI isn't on truly high-value work yet. When the return is high enough, the input cost is worth ignoring.

[Mark Li @iammutex]

@ekcheungAI Under the hood that video is frontend code. A small tweak is a small diff, so a revision shouldn't cost another 4% every time. The first build is the expensive part.

[Mark Li @iammutex]

@HarryTandy GTM. Code is in the training data, GTM data isn't: who works where, their email, what the company raised last week. Claude can write the email but can't know who to send it to. Let your agent get that through @Glasserai 😎

[Mark Li @iammutex]

Semrush subscriptions are dead.
My agent just reverse-engineered Manus's entire 2026 Google Ads strategy.

20 countries. One prompt. $0.24.

What it found:

→ Their biggest bet was Brazil, not the US
→ May alone: 513K paid visits
→ The ad that won never said "free." It said "AI that does the work."

One month of Semrush ($99) pays for 412 of these runs.

No seat. No subscription. Pay per call.

Steal my setup: https://t.co/JRrBd6PBOr

[Mark Li @iammutex]

@Mr_Buzzoni Isn't this common sense? What really makes the difference is which tools the agent can reach. Connect your inbox and it handles your email. Connect your site and it writes the blog. Connect @Glasserai and it gets verified emails, company and funding data.

[Mark Li @iammutex]

@pauliusztin_ Compaction is a lossy save. I'd never let the summary be the only copy: have the agent write the root cause and what it already tried to a file before compaction, then reread it after. Summaries drop details, files don't.

[Mark Li @iammutex]

@huesofbanter @cursor_ai I don't think notifications are what fixes this. Agents drift because each one holds its own context. The real question is whether they share context and memory by default, so a plan change reaches every agent without anyone telling them.

[Mark Li @iammutex]

@vallver That's a lawyer, not a brainstorm. Once it picks a side, the rest of the answer is the closing argument. I'd make it stop after the list, and let a second pass try to kill the first pick.

[Mark Li @iammutex]

@alexhvnsen We tried the interview step too. It pulled out things we'd never have thought to say up front. Hermes as the one who asks first is a really good call.

[Mark Li @iammutex]

@DottChen I'd only pay a markup for something the model can't know. Translation it already knows, so next-intl plus your own key wins. A service earns its 10x when it brings fresh data the model can't produce, not when it resells the same tokens with lock-in.

[Mark Li @iammutex]

@neil_xbt > pay the expensive model for the thinking, pay the cheap one for the typing

It's a factory: the engineer draws it once, the line builds it a thousand times. The catch is the drawing. If the plan is too loose for the cheap model to follow, you pay $40 to fix $10 of typing.

[Mark Li @iammutex]

@stretchcloud Most of what agents scrape is already for sale as clean records. Having your agent crawl it is like building your own crawler after Google exists. Pay cents per record and save the context for thinking. Let your agent get it through @Glasserai 😎

[Mark Li @iammutex]

@ericosiu Agreed. The subscriptions from model companies are subsidizing tokens by dozens of times right now. That's an unfair advantage the labs have over every other product built on their models.

[Mark Li @iammutex]

@stretchcloud Catching the conflict is the easy half. The hard half is who wins. Each agent is guarding a different person's requirements, so two agents shouldn't negotiate it away between themselves. They should escalate to whoever owns that decision.

[Mark Li @iammutex]

@pggalaviz @jullerino Desktop has the same bottleneck, we just hide it behind a big screen. Talk is the easy part. The product is a review surface that lets you reject a 40-file diff in 30 seconds, not a microphone.