Llama Ventures lets your AI make the pitch.

A founder looking for venture money usually starts with a name. Someone who knows someone. Someone willing to forward a deck. Llama Ventures offers another starting point: a command you can hand to an AI agent. On its public pitch page, the instruction is disarmingly ordinary. Run the tool. Send the pitch and deck. Receive a structured assessment. An encounter normally arranged through a social network has acquired a software interface.

• The business: early-stage capital for AI-native companies.
• The useful extra: public diligence frameworks founders can use before pitching.
• The tension: automate the paperwork while keeping responsibility for judgment human.

That small detail is a useful way into Llama, a Silicon Valley venture firm whose publicly identified founder, Jack Feng, appears alongside investors and operators including Herman Zhou. The firm backs businesses across AI infrastructure and industry applications. But the more interesting question is operational: when a fund invests in AI, how much of its own work should change?

A pitch with an install command

Llama’s public agent page advertises compatibility with tools that can execute shell commands. It promises an assessment across twelve dimensions, including team, market, defensibility, traction, and the fundraising ask. The external pitch route is described as requiring neither an account nor a token, with rate limits tied to email and IP address.

For a founder, the attraction is concrete. Information already assembled in an agent’s workspace can become the starting material for a pitch. The advertised assessment supplies categories against which to examine the business. An evaluation still has to become a conversation, diligence, and an investment decision. Those are separate steps, with separate stakes.

The questions are part of the product

Llama also publishes Llama OS Public Skills, a collection of reusable frameworks for AI-assisted investment work. They cover diligence, technical assessment, market mapping, and investment-committee reports. A founder can apply the same questions to their own startup. An investor can use them to structure research. The repository describes this public methodology as one layer of a larger private operating system.

There is an appealing commercial logic here. Publishing questions makes the firm’s thinking available before a relationship exists. It also lets prospective founders discover how their case might look from the other side of the table. The reusable asset is a framework; the difficult work is supplying evidence and deciding what it means.

Zhou’s October 2025 startup checklist makes that distinction tangible. At the idea stage, he proposes eight to twelve weeks and twenty to thirty serious conversations. Identify who discovers the problem, who decides, who pays, and who uses the result. Set a failure line in advance. If the expected signals fail to appear, change direction.

“They mistake enthusiasm for validation.”

Herman Zhou, The Early-Stage Startup Checklist

The advice gives a founder something copyable: a deadline, a research task, and a rule for reconsidering. Before building more, define what would justify building more. It is also a test of temperament. A failure criterion serves little purpose if the team quietly moves it whenever the answer becomes uncomfortable.

A human opinion needs a human

The public software record offers a less polished, more revealing companion to that philosophy. In August 2026, Llama’s CLI changelog described a problem with contested facts: once challenged, a true claim could remain excluded from generated summaries even after its accuracy was confirmed. A new operation allowed the contest to be lifted while preserving its history.

That is a small failure with a large implication. A research system needs a way to recover from mistaken skepticism. Otherwise, yesterday’s objection becomes tomorrow’s invisible filter. The changelog does not establish an investment loss; it documents a defect in how information could move through the workflow.

September changes drew another boundary. Entries representing a human’s subjective view had to identify the speaker and preserve their actual words. An agent summary would appear alongside that original wording. This is a practical response to a familiar risk: fluent prose can make an interpretation look like a statement someone actually made.

The lesson travels beyond investing. Keep the underlying statement attached to the summary. Give corrections a path back into the record. Automation becomes easier to inspect when it leaves enough evidence for someone else to disagree.

From a bid sheet to a living world

Llama operates both a direct-investing fund and a fund-of-funds business. Its homepage puts typical startup checks at $500,000 to $5 million. For founders comparing it with another seed investor, the practical question is whether its technical judgment and operating support fit the work ahead. Llama advertises help with product, hiring, market entry, and later fundraising.

The investment portfolio brings the argument back to businesses. In September 2026, Quotr announced a $4 million seed round led by Llama. Its software works on construction takeoffs, estimates, proposals, and procurement. The customer problem is wonderfully unglamorous: contractors need to turn drawings into a bid before the opportunity disappears.

Quotr said the funding would expand enterprise deployments, its supplier network, and access to the platform. This is AI attached to a recognizable workflow, buyer, and budget. The $4 million is the announced round total; it should not be read as Llama’s individual contribution.

That same month, Visko announced a $10 million pre-seed round led by Llama and opened access to Orbis, its live world model. Visko describes continuous interactive video rather than a finished clip. Its proposed applications reach into simulation, robotics, and creative media. Put the two investments together and Llama’s range becomes visible: a contractor’s estimate at one end, a research-heavy model at the other.

Who gets paid when agents get busy?

Llama’s infrastructure writing helps explain the connection. Gavin Zhu argues that agents increase demand for cloud services while pushing suppliers to prove their value. In his own example, Vercel’s $20-per-person monthly team pricing helped prompt a move to Cloudflare. AI assistance made the extra work tolerable for his small workload. He also draws a limit: production systems bring reliability and operational demands that a few generated scripts cannot erase.

For founders, that is the useful challenge in Llama’s approach. Name the customer. Measure the friction. Show why the next layer of software earns its place. More activity can enlarge a market while competition squeezes the businesses inside it. A persuasive AI pitch has to explain who keeps the value.