Synthesis
Making a business agent-ready
Letting other people's agents read, use and buy from your business, and deciding deliberately whether to block them.
Four episodes, two doing the heavy lifting. Greg Isenberg, as host, argues an agent is ‘a customer with a credit card and no eyes’: facts must be in plain text, buying flows tested with several agents, bot protection audited, and blocking paired with a plan to let agents in safely. His examples (Amazon, Delta, Yelp) and a nine-stage prediction modelled on banks versus Mint and Plaid are his own claims and forecasts. Parag Agrawal, founder of an agent-web infrastructure company, agrees on design: ship a machine-readable version, copy-paste agent prompts in docs, and keep the human to authorising and paying. He notes that declaring agents currently increases blocking, so safe-looking practice degrades products, and that agent traffic statistics are inflated by crawling. Both speakers sell or promote related offerings, so predictions on ads and attention monetisation are opinion. For people running their own agents, the lesson is the other side: expect friction and inherit only the access the site will allow.
Key points
- Put the facts agents need in plain text on the page, not in video, PDFs or last checkout screens — 1 episode. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 06:07)
- Test your own buying flow with several agents and fix where they get stuck — 1 episode. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 22:19)
- Audit bot defences deliberately; some blocking is legitimate, some accidental — 2 episodes. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 22:43) (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 33:01)
- Design onboarding and docs for a person working with a coding agent; the human authorises and pays — 1 episode. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 47:08)
- Declaring good agents currently causes more blocking; incentives are unaligned — 1 episode. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 37:29)
- Agent-friendly flows win repeat agent choice — 1 episode. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 19:31)
- Advertising and attention models are predicted to weaken as agents become users — 2 episodes. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 29:29) (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 20:28)
Where guests disagree
Should businesses block agents? ⚠️
- Block, or accept only on paid terms, protecting ads and the customer relationship (Amazon, Delta, Yelp as described by the host) (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 02:18)
- Welcome agents and design for them; blocking mostly protects a habit (host’s view, Parag’s design stance) (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 28:39) (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 47:08)
Frameworks
- Nine stages of the war on agents — The host’s prediction of how sites and agent platforms settle their conflict, modelled on banks versus Mint and Plaid. The host does not describe stage 6 aloud. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 18:23)
-
- Companies block agents because agents skip ads and upsells (today)
-
- Agents act more like humans to get through: an arms race of blocks and workarounds
-
- Market splits: sales businesses let agents in, data businesses charge for access, giants build their own agents
-
- A trusted middle layer for agent identity and payments becomes the standard way in
-
- Companies that welcome agents win customers from those that block them
-
- (not described in the episode)
-
- Everyone competes for agents: better prices, faster checkouts, pages written for agents
-
- Ads return: companies pay to be the first thing an agent recommends
-
- Three camps on agent access — How companies sort by how they make money: block, accidentally block, or welcome. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 04:43)
- Block camp (e.g. Amazon, Delta, Yelp for non-payers)
- Accidentally-block camp (e.g. Zillow, Walmart)
- Welcome camp (e.g. Meta, Shopify, Visa and Mastercard)
Tactics
- Run an agent test on your own buying flow — Pull 10 real buying questions from the support inbox, give each to two or three different agents, and watch where each gets stuck. Do the same on a competitor’s site. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 22:19)
- Audit whether you are blocking agents by accident — Review bot protection set up years ago, then decide deliberately what to allow. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 22:43)
- Make the offer readable by agents — State total price, delivery date, what it works with and how returns work in plain text on the page, not in a PDF or on the last checkout screen. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 23:22)
- Make your advantage readable too — If you cost more but deliver faster, show the delivery date so an agent comparing options can see why you are the right answer. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 23:45)
- Make a machine-readable version of your site and put copy-paste agent prompts in your docs. — Parallel has a human version of its site with a toggle to a machine version; docs say ‘put this prompt into your agent’; the human only authorises, adds a card and accepts terms. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 46:07)
Stories & examples
- Banks, Mint and Plaid: Banks fought Mint.com until Plaid built a proper connection between apps and banks and became worth billions; banks moved from blocking to signing deals to charging for access. Lesson: Access conflicts tend to settle through a trusted intermediary and paid access rather than permanent blocking. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 17:31)
- Two airlines: If one airline lets your agent book in seconds and another makes you book yourself, you know which your agent will pick next time. Lesson: Agent-friendly flows win repeat agent choice. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 19:31)
- Blockers building their own agents: Amazon (Rufus) and Delta (an AI concierge) block outside agents while building their own. Lesson: The real position is that agents are fine if the company owns them, which tests whether ‘owning the customer relationship’ is serving the customer. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 11:37)
Numbers
- Estimated agent usage as share of human web usage: 15 % (10 to 20) (self-reported, approx.) (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 06:50)
- Remaining growth in agent web usage: 7500 x (5,000 to 10,000) (self-reported, approx.) (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 06:59)
- Estimated agent web usage versus humans, as predicted three years ago: 1000 x (self-reported) (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 02:27)
- Real buying questions to test with agents: 10 questions (host’s framing) (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 22:19)
Caveats & warnings
- The stages may not happen in a neat order; some companies are at stage six while others are at stage one. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 21:39)
- Some blocking is legitimate: much traffic calling itself an agent is a scraper, and some blocks are security systems working as built. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 09:27)
- Avoid businesses whose edge is a trick to get mentioned by AI; tricks stop working. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 27:48)
- Agents will not make every market a race to the cheapest, since the people behind them also want reliability and speed. (presenter · The 9 Stages of AI Agent Adopt (Oct 2026) ▶ 23:48)
- Declaring agents as agents currently causes more blocking, so the safe-looking practice degrades products until incentives are aligned. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 39:00)
- Agent traffic statistics include crawling for pre-training, so raw ‘agent traffic exceeds human traffic’ overstates agent use for equivalent work. (Parag Agrawal · Parag Agrawal (Oct 2026) ▶ 06:29)
How much to trust this
- Evidence: 4 episodes, 2 independent interviews (each video counts once).