Case 11 · Graduation Day
Smithers
The rare pitch where the Q&A outperforms the pitch — the benchmark answer earned "most honest and elaborate answer I've heard so far" from the floor — but the prepared minutes open with a false start on the company's own name, disclaim the moat on stage, and end with no number, no buyer, and no ask.
01 Scorecard · 12 dimensions
02 Klaff 5-lens audit · 2.5/5
THE ONE REORDER — Memorize sentence one — the identity line — so the restart can never happen, and replace "come talk to me" with an ask. The middle already works.
03 Delivery
144 fillers ("like" ×76, "um" ×28, "uh" ×12). "Like" alone is 3.4/100 and clusters in the Q&A answers — the prepared portion is noticeably cleaner. Q&A rehearsal is the drill.
04 Beat-by-beat
| Time | Beat | What works | The fix |
|---|---|---|---|
| 0:00 | The restart | Honesty about nerves humanizes | "Workflow Ro-- ... Sorry, let me start over" — on the company's own name. Memorize sentence one; never open cold. |
| 0:35 | Surprise-signal story | "Three paragraph open letters... begging me to fix all the bugs" — the best traction color in the pitch | Attach counts: N issues in week one, N integrations, N letters. |
| 1:15 | Problem | "The worst thing that can happen is things go halfway right" — quotable, true | Add the victim: whose 2 a.m. is ruined, what it costs them. |
| 1:45 | Evals & the bug promise | "Over six thousand evals" + first-prompt-failure-is-a-bug — a real differentiator stated as a promise | Say what an eval buys the listener: measurable reliability, not vibes. |
| 2:30 | Live rate-limit save | Accidental live demo: Smithers parks itself and resumes when the rate limit clears | Make it deliberate — script the failure, let the room watch the recovery. This is the demo. |
| 3:00 | Time travel | "See every single frame... rewind and time travel" — the trust argument | Connect it to a buyer: who needs frame-level audit of an agent workflow? That's the enterprise wedge. |
| 4:15 | Vision | "The first Smithers workflow that just runs forever indefinitely" — vivid, memorable | Keep verbatim. |
| 4:45 | Close | The coward line wakes the room up | "Moat is doing the last mile for enterprise" needs one sentence of what that is — then an ask, not "come talk to me." |
| 5:40 | Q&A: benchmarks | Names Anthropic, concedes the speed gap, promises public benchmarks — complimented from the floor | Move this honesty into the pitch. It's his highest-status material. |
| 9:30 | Q&A: customer story | "Use Smithers to rewrite Smithers in Python" — funny, sticky, proves meta-capability | Anonymous and unverifiable as told. Get one nameable production user. [confirm] |
| 11:00 | Q&A: EVM story | Hill-climbing methodology, "fastest EVM ever in terms of benchmarks" [confirm] — founder competence on display | This is the missing Team slide. One line of it belongs in the pitch proper. |
05 The gaps investors block on
- No ask. "Come talk to me" gives the room nothing to act on — no amount, no milestone, no pilot slot.
- No business model. Open-source-forever plus an unexplained "last mile for enterprise" is a stance, not a revenue line.
- Traction is anecdote, not number — issues, integrations, and open letters with no counts; the best customer story is anonymous.
- The moat was conceded on stage ("no such thing as a software moat", Anthropic parity in six months) with the enterprise counterweight scattered across Q&A instead of built into an argument.
- No one-sentence identity — a first-time listener cannot repeat what Smithers is, and the pitch opened by getting its own name wrong.
Name the enterprise product. The fork is open-source project vs. venture company. Distribution is chosen ("open source forever"); the company is one clause long ("last mile for enterprise"). Take the three deltas he rattled off against Anthropic — multi-machine, multi-agent, frame-level audit/time travel — and make them the paid tier, with one design-partner logo [confirm]. The ingredients are already in his mouth; durability, observability, and trust are exactly what enterprises pay for and hyperscaler harnesses skip.
“Smithers is the open-source runtime that makes agent-built workflows durable by default — free forever for builders, paid for the enterprise last mile: multi-machine, multi-agent, every frame auditable.”
06 The rewritten pitch
I'm Fakhri, and I build Smithers — the open-source runtime that makes agent-built workflows durable by default. When I shipped it I expected silence; side projects die to quiet churn. Instead, people opened issues on day one, built integrations I never asked for, and wrote me three-paragraph open letters begging me to fix bugs. That's what a real problem looks like: when a workflow goes halfway right, a human has to crawl in at 2 a.m. and clean up the mess.
With Smithers you type what you want — any harness, any model — and the agent one-shots a working workflow with its own UI. We hold that promise with over six thousand evals, and if your first prompt doesn't produce a working workflow, we call that a bug and it joins the eval set. Watch what happens when I get rate limited — the workflow parks itself, waits, and resumes. Every frame is recorded; you can rewind and time-travel to any point in the run.
Anthropic added workflows to their harness [confirm: six months] after we shipped — what they shipped is our MVP, and I expect feature parity on the basics within six months. The features aren't the moat. The last mile is: workflows across many machines, across many agents, with frame-level audit — the tier enterprises pay for while the runtime stays open source forever.
Today [confirm: N] people run Smithers, including [confirm: one nameable production user + what they run]. I'm the person who [confirm: one-line EVM credential, if benchmarkable]. I'm raising [confirm: $X] to land [confirm: N] enterprise design partners and publish the public benchmarks. If your agents build workflows that break while you sleep — come break mine.
Brackets are the founder's homework — the copy is done, the proof is theirs to supply. Every claim must match what's actually built.
07 Do these before the next take
- Write the ask — amount, use of funds, milestones. One paragraph of work; currently zero words.
- Define the enterprise last mile as a product — the three deltas (multi-machine, multi-agent, frame audit), a price shape, and one design-partner conversation [confirm].
- Count the traction — issues, integrations, users, letters. The color is already great; attach numbers, and land one nameable production user.
- Memorize sentence one — the identity line — so the restart can never recur, and cut "sorry/nervous" from the open.
- Publish the benchmarks promised in Q&A, then move that honesty into the pitch proper — it's his highest-status material.
- Filler drill on Q&A, not the script — "like" ×76 lives almost entirely in the answers. Rehearse answering the room-prep questions aloud.