AgentSee Podcast

Abrar builds GradeDrive, an AI marking tool for schools, with a non-technical co-founder, and helps a second startup where the product team ships its own ideas. He explains what makes that safe: a skill that gives every idea its own worktree, database and dev server, a read-only environment the agent can be trusted with, guidelines written after a feature that ran a whole class's PDF in the browser, and a draft pull request that spins up an environment of its own. James asks about native apps, things going off the rails and what the job is now. The answers: web is easy and native is not, it is easier than ever to build sloppy stuff, taste still comes from people, and Abrar has not opened a code editor in six months.

Chapters:

• 00:00 GradeDrive, and another startup
• 01:36 How it used to work: idea, engineer, ticket, wait
• 04:10 Vibe coding: quick, motivating, throwaway
• 07:15 Prototypes that carry the intent
• 10:28 Enablement: a read-only environment, a worktree per idea
• 14:43 When a vibed feature went wrong
• 19:00 Agents that check their own work in a browser
• 21:42 Web is easy, native is not
• 26:25 One-shotting a feature, and the draft pull request
• 30:06 Easier than ever to build sloppy stuff
• 34:43 Engineering is enablement; then gravitate to product
• 37:40 Coding is an abstraction; taste is still ours
• 42:11 The role is exhausting: the thinking comes first
• 47:34 Where people hit the wall
• 52:10 Outro

Mentioned:

• GradeDrive, AI marking for schools
• Claude Code, and the Claude desktop app the co-founder uses
• git worktree, one checkout per idea
• Draft pull requests on GitHub
• The Ralph Wiggum loop
• Expo, the native app toolchain mentioned

Of the questions on the site, this one touches:

• 3. What is a product manager when the specification is executable?
• 5. How do you run acceptance testing on something produced faster than a person can read it?
• 7. What does taste mean when volume is free?
• 8. Do engineers still want the job? What is it now — author, reviewer, director, or something we don't yet have a word for?

AgentSee is two people building a company in public, with agents doing the work.

https://agentsee.work · [email protected]

Subscribe: https://agentsee.work/podcast/feed.xml

James set up AgentSee's own mail server on a Swiss VPS, with an agent writing the OpenTofu and most of the configuration: two afternoons' work rather than the weeks it would once have taken. Abrar asks what was hard, and the answer is the click ops inside Stalwart's admin panel, a CLI that was not built for an agent, buttons the model insisted existed, and making the result reproducible afterwards with a snapshot script. They also cover borrowing sender reputation, the monitoring that is not built yet, whether either of them would recommend this to a client (no), and what it is like to let an agent drive the browser.

Chapters:

• 00:00 Why mail at all
• 02:15 The alternatives, and why Infomaniak
• 05:56 OpenTofu, and what the agent wrote
• 07:26 Stalwart: the click ops and a clunky CLI
• 12:31 Buttons that don't exist, and the snapshot
• 16:40 Set and forget: spam, SES and the relay
• 19:00 Monitoring, and what we have not built
• 21:56 Reputation: borrow someone else's
• 23:51 What the server holds
• 25:48 Would we do it again?
• 29:53 Letting an agent click
• 36:53 Where to find it all
• 38:00 Outro

Mentioned:

• Issue No. 2: the mail server, written up
• The infrastructure, as OpenTofu, on GitHub
• Stalwart mail server
• OpenTofu
• Infomaniak Public Cloud
• SMTP2GO

Of the questions on the site, this one touches:

• 4. Who is accountable for a system that nobody typed?
• 8. Do engineers still want the job? What is it now — author, reviewer, director, or something we don't yet have a word for?

AgentSee is two people building a company in public, with agents doing the work.

https://agentsee.work · [email protected]

Subscribe: https://agentsee.work/podcast/feed.xml

James Hartt and Abrar Mahmood introduce themselves and the premise of AgentSee: a software agency for a world where working software can be prompted into existence. Agencies were organised around protecting developer time and billing for it; when that is no longer the expensive part, what does a client pay for? Their answer, for now, is the last stretch to production, which is as hard as it ever was.

Chapters:

• 00:00 Welcome, and who we are
• 02:32 The premise: what agencies were built to protect
• 04:43 What a client pays for now
• 05:49 Where the value is: the last five percent
• 07:34 Wrap up
• 07:59 Outro

Mentioned:

• Issue No. 1: the premise, and the ten questions

Of the questions on the site, this one touches:

• 1. If writing the code is no longer the expensive part, what is the client actually buying?
• 7. What does taste mean when volume is free?

AgentSee is two people building a company in public, with agents doing the work.

https://agentsee.work · [email protected]

Subscribe: https://agentsee.work/podcast/feed.xml