Amp is the New 37signals

Justin introduces Amp, which he describes as a cloud “meta harness” for LLM coding. Each chat session gets a virtual machine, allowing his Rails app to run on fast cloud previews that he can share with teammates. Amp’s “portals” expose a normally local development environment through a browser, while a small widget lets him select an interface element and request a quick change. The panel notes that this makes it easy to multitask with agents, resume work from a phone or laptop, and keep experiments in a sandbox separated from the rest of a developer’s computer.

Justin says Amp is interesting not only as a tool but as a company with a philosophy, comparing it to 37signals in the early Basecamp era. Amp’s 21-person team is aligned around two simple goals: build the best product and make the numbers go up and to the right. Its podcast, “Raising an Agent,” emphasizes removing friction: no mandatory pull requests or human code review, no quarterly or annual goals, and only one short Monday meeting where people set their own weekly goals. Team members are expected to do work that gives them energy while remaining focused on customers.

Jordan highlights that Amp’s leaders and employees visibly use their own product; feature requests and bugs are often answered directly by the CEO or CTO. Justin explains that all 21 people can work across the product rather than belonging to departments: they do marketing, customer service, sales calls, and client visits. If one person cannot attend a client meeting, another can go because ownership is shared. The panel contrasts this with traditional software teams, where engineering, product, design, sales, and leadership form separate factions and private chats often become places to complain about other departments.

Amp’s model prompts a discussion about pull requests, specs, QA, stand-ups, agile points, and project-management tools. Justin recalls how the old conveyor belt felt reassuring when code was fragile and engineering was expensive, but says much of it may have been “cruft” or process designed to make people feel busy and give jobs a defined purpose. Brian cautions that the label matters less than the substance: a PR, PRD, spec, prompt, Slack exchange, or meeting can all work if they produce good thinking and collaboration. The panel’s conclusion is to remove unnecessary friction without treating any particular tool or ritual as sacred.

Justin says Transistor initially considered replacing Shortcut because team members disliked the tool. At the retreat, the five-person team asked what project management was actually supposed to accomplish, considered Trello, briefly began vibe-coding a replacement, and finally returned to Shortcut. They realized the issue was not the software but the process, expectations, impatience, personal preferences, and habits carried over from previous jobs. Brian says his own company has naturally become less structured: two engineers prefer high autonomy, and product manager Jess moved from enforcing rigid processes for a payments product to asking for more tools and access.

The panel argues that companies now need to reconsider who they hire and how teams are organized. Justin believes AI-native, hungry younger employees may have an advantage over people attached to old workflows, quoting DHH’s claim that the age of agents will be hard on “cantankerous” programmers, designers, managers, and product managers. He also says professionals should have access to tools such as Claude or Codex, and suggests that companies still using only Copilot may become acquisition targets. Brian adds that planning and risk management mattered when engineering was expensive, whereas the new pressure is speed.

Justin describes a thought exercise about Fiverr: roughly $300 million in market capitalization, $400 million in revenue, $200 million in labor expense, $150 million in marketing, and $200 million in cash. He calculates that the market is effectively valuing the operating business at about $116 million after accounting for cash. The panel wonders what happens when an activist-style leader asks why a company needs hundreds of employees if a small, AI-enabled team can produce similar output. They do not resolve the question, but treat the gap between legacy organizational cost and agent-enabled leverage as a major opportunity.

Brian unveils Homegrown Systems, a productized service whose headline is “run marketing in-house with agents that you own.” For an initial fee he plans to install a team including a chief of staff, analytics reporter, SEO and AI-visibility optimizer, content writer, and strategist for marketing experiments. The agents can audit a site, write competitor pages, produce reports, request approval, and deliver updates through email, Slack, or text. Customers retain their own cloud, model accounts, data, company context, brand voice, positioning, and analytics connections rather than outsourcing everything to a black box.

Jordan asks what a customer practically receives for about $2,000. Brian says it includes agent skills, documentation, process design, company-specific training, integrations with systems such as Google Analytics, Stripe, HubSpot, Fathom, or Plausible, and an onboarding period in which the company’s market and voice are learned. The panel distinguishes this from downloading generalized open-source skills such as Corey Haines’s marketing collection: those may be a useful first 5%, but they are not trained on a particular company or its preferred communication style. Brian is still deciding whether the core product is software, service, or both.

Brian says his first validation step was mentioning the unnamed idea in a newsletter and inviting replies; several people responded, leading to early customer conversations. He plans to approach Builder Methods members, industry contacts, and other concentric circles before seeking roughly ten initial clients. The group compares the offer to a recruiting service or marketplace for agents, while Brian favors a simple core package with optional agents and support. They connect this to Joel Gascoigne’s observation that AI makes hesitation expensive: once a decision is made, shipping is fast. Justin says Transistor increasingly sees hiring as an experiment, preferring a $2,000–$3,000 agent-enabled test to committing to a $100,000 employee role.