Making $$ with AI Agents

Business Opportunity & Market View (AI Agents)

Sequoia adoption snapshot (domains where agents are used)

• Software engineering: ~50%
• Back office: ~9%
• Marketing & “copyrighting”: ~4%
• Sales & CRM: ~4.3%
• (Others not fully legible in the transcript.)

Core reaction / thesis

• Agent use is underapplied relative to how many workflows are ready to be disrupted.
• Even software engineering may be overestimated; the speaker argues “frontier agentic development” is accelerating beyond “autocomplete/coding assistant” (e.g., IDE-less, multi-instance, autonomous PR/agent loops).

---

Timing, Breakthrough, and “Why Now”

• Agents reached a meaningful autonomy breakthrough roughly “4–5 months ago” (speaker references Opus 4.5 as a turning point).
• Models are considered smart enough now to execute multi-step tasks with tools/context.
• Remaining barrier is deployment across roles.

---

TAM Framing (High Level)

• Agent TAM is framed as potentially enormous:
  • Not just “$1T” (Sequoia number), but possibly “tens of trillions” by mapping to the value of all white-collar labor (GDP-scale framing).
• The speaker emphasizes many people don’t fully “get it” until they use agents hands-on (not one-shot queries).

---

Business Execution Playbooks & Frameworks Mentioned

“Agent-first” business logic (managerial system, not just tools)

• Shift from too sporadic experimentation → to daily, calendarized use.
• Treat agent setup like management 101, but with automated quality control:
  • You need “checks and balances” as you scale from 1 agent to fleets.

Skill + rubric system (HyperAgent primitives)

• Skills
  • Reusable “job playbooks” you teach/compose so agents reliably perform a role.
• Rubrics
  • Evaluation/quality scoring frameworks (using an LLM judge) to create an eval loop over time.

Evaluation loop (scoring + iteration)

• Run agent tasks → score output with rubric (judge-as-a-service) → inspect failures → update skill/guidelines → re-run.
• Rubrics enable trend lines and continuous improvement without constant human review.

Cost framing (unit economics mindset)

• Reframe agent costs vs human opportunity cost:
  • Even if tokens cost (example given) like ~$150 for a board memo draft, the comparison is the time/value of a human producing it, not “software subscription pricing.”

---

Product Walkthrough: HyperAgent (Strategy + Differentiation)

Positioning

• Compared to other agent builder products:
  • “OpenClaw-like” products = more Linux / less turnkey UX
  • HyperAgent = Mac-like UX: secure, cloud-native, visual, intuitive
• Competitive edge claims:
  • More powerful tools out of the box
  • Better UX and deployability
  • Designed to support “fleet” usage (multiple agents/roles) + oversight at scale

What HyperAgent can do (concrete examples)

• Research → app generation → V1 build end-to-end
  • Example startup idea: hyper-local real estate market reports
    • Agent researched the market, produced a research brief, did competitive analysis, validated demand via Reddit, then built a version 1 product (because it includes coding capability).
• Tool integrations / capabilities (conceptual demos):
  • Connect to apps/accounts like Slack, Gmail, Notion, “Granola,” etc.
  • Pull in Twitter
  • Use Google Maps / Street View to find imagery (billboard points of interest)
  • Use images as seeds for AI image/video generation
  • Redesign a house using interior/exterior references (e.g., from Zillow imagery)

“Command center”

• Manage an overview of many agents (e.g., content marketer, market researcher, email responder).
• Deploy agents into team workflows (e.g., always-on Slack agent behavior).

---

Skills & Onboarding (Actionable How-To)

Skills: how to think about building them

• Analogy: the base model can be “smart generally,” but skills provide the playbook for a specific role.
• Example shown: creating a “Greg Eisenberg” style drafting skill (with guardrails like don’t impersonate).
• Process includes:
  • Research the person/style
  • Distill into a structured skill
  • Set scope (e.g., drafts only, no auto-posting)
  • Define topics (e.g., contrarian AI takes, solopreneur/bootstrapping themes)

Evergreen improvement loop

• Skills shouldn’t be one-shot. After each use:
  • the agent learns via accumulating learnings/memory, and/or
  • human feedback corrects outputs
• Strong encouragement: don’t quit after first try—expect a “messy middle” before high-quality output.

Rubrics: scaling quality without constant human review

• Rubric example:
  • “Score Greg-style content” on dimensions the user cares about
  • Use an LLM judge; view trend lines of performance
• Practical cost lever:
  • If rubric scores stay high, reduce model quality/cost (example mentioned: dropping model tier to cut cost ~5x with minimal score impact).

---

Scheduling + Adoption Targets (Behavioral KPI)

• Advice for builders: run HyperAgent every day for a commitment window:
  • 30 / 60 / 90 days
• Mechanism:
  • Put it in the calendar (e.g., 30 minutes blocks) → compounds learning and outputs.
• Milestones mentioned for solopreneurs:
  • First internet dollar (re-anchors confidence)
  • $10K/month as a major inflection point (often leads to quitting job / going all-in)

---

Marketing/Sales Motion & Growth Strategy Ideas (Agent-First GTM)

• Two major angles for agent-driven businesses:
  1. PLG wedge (product-led growth):
     • Let users try AI-powered workflows
     • Example category: “PLG agent” products like the speaker mentions (e.g., OpenClaw as a concept)
  2. Top-down “pay for risk mitigation” (enterprise consulting style):
     • Pitch to CEOs/boards: “We’ll fix your AI problem” for large checks
     • Speaker cites $100M+ check framing
• “Agent command center” as the future organizational artifact:
  • Treat agents as role-mapped coworkers orchestrated in one place

---

Metrics & KPIs Explicitly Referenced

AI usage / adoption metrics (Sequoia charts)

• % agent deployment by domain:
  • Software engineering ~50%
  • Back office ~9%
  • Marketing/copywriting ~4%
  • Sales/CRM ~4.3%

Profitability/scale metrics (company-level, illustrative)

• Airtable sponsorship claims:
  • ~half a billion in revenue
  • ~$1B in bank
  • “generate $100M of free cash flow this year” (claimed)

Token/cost proxy examples

• Board memo generation: ~$150 tokens (used to illustrate opportunity-cost framing)

Growth/adoption milestones

• $1 stranger → brain rewired
• $10K/month → likely quit job / go all-in
• Break-even described qualitatively as confidence milestones (no exact numeric target given)

---

Concrete Recommendations (Actionable Takeaways)

• Don’t do light experimentation—use frontier agents with ambitious tasks for full autonomy.
• Commit daily to agent building (30–90 days) and make it part of your workflow.
• Build “skills” like playbooks, then measure outputs with rubrics for scalable quality.
• Design for fleets (command center + oversight) rather than only one-off agent threads.
• Expect a ramp: improve through iteration; don’t abandon after first “meh” outputs.
• Choose a real business problem first (the hardest part isn’t the UI; it’s selecting a valuable opportunity). HyperAgent onboarding is described as helping infer use cases from your existing context.

---

Presenter / Sources

• Presenter / guest: Howie Lou (co-founder and CEO of Airtable), discussing HyperAgent (hyperagent.com)
• Host / interviewer: Greg Eisenberg (referred to throughout)
• External source referenced: Sequoia (charts on agent deployment and “$1T up for grabs” framing)
• Model/reference points mentioned: Opus (including “4.5/4.6/7” tiers), GPT-5.4 (spelled “GB 5.4” in subtitles), Claude Sonnet (used in rubric cost example), GitHub Copilot, and Andre Karpathy (discussion referenced)