Video summary

I let GPT-6 Astra run my business… it’s insane

Main summary

Key takeaways

Technology

Technological concepts / product features discussed

  • Agentic “agency” workflow (promptless UX): Instead of users prompting an AI, the AI continuously suggests actions. Users mainly approve/reject with fast yes/no interactions (rather than swiping or long prompts).

  • Always-on multi-agent automation: The system runs many sub-agents in parallel to handle workflows like:

    • Customer support ticket triage and responses
    • Growth tasks (e.g., drafting posts/answers based on social mentions)
    • Automated code changes and PR creation/merging
    • Integrations/testing/issuing tasks to maintainers

  • Connected tooling across the stack: Agents monitor and operate on:

    • Cloud resources
    • Repositories (GitHub)
    • Support systems/tickets
    • Documentation
    • Communication channels (e.g., email/Slack)

Agents can also:

- Create **PRs**
- Run **local tests**
- Publish/merge after validation

“Browser use” is mentioned as a way to automate interactive web tasks, including generating and embedding outputs into repos/README.

  • High trust automation (with human-in-the-loop possible):

    • The speaker describes trusting agent-suggested changes enough to publish PRs and merge, while remaining responsible for higher-level constraints (e.g., code execution credits/subscriptions).

  • Resource scaling and operational constraints:

    • Running many browser sub-agents on a local Chrome setup can crash due to memory limits (e.g., laptop/Chrome out-of-memory).
    • Proposed workaround: move execution into cloud browsers and use profile sync there.
    • Another constraint: agents at scale can burn API/execution credits quickly.

  • “Alignment problem” framing:

    • The main requirement isn’t just “smarts,” but alignment—ensuring the agent’s notion of “goal completion” matches the human’s taste, constraints, and what not to do.
    • The user’s “judgment” is treated as the remaining scarce resource, especially deciding which tasks are worth doing and at what cost.

  • Capturing personal preferences into machine-readable form:

    • Input can be reduced to a single “dreams” markdown file.
    • Agents then propose tasks and iterate approval.
    • A future “alignment” step is described: generate preference data from past interactions (e.g., analyze historical sessions, Slack/Gmail/Calendar, code reviews, and “ugly/stupid” feedback) into a markdown/JS-like preference representation.

  • Created-content instead of ranked-only recommendations:

    • Future recommenders are described as agents that create new actions/results, not just rank items (unlike Netflix/Google/Tinder).
    • The “next thing” isn’t only selected from a list; the system may generate tickets, PRs, or even software on the fly.

  • Automation beyond support: commerce and marketplaces disruption:

    • Agents will perform purchase booking decisions (e.g., find the best train route/cheapest flights, then book).
    • This suggests marketplaces/platforms could become less central, since agents connect buyer/seller intent directly (“dream-to-match economy”).
    • Companies must be findable and have good reputation/review data, since agents will choose based on that—not just marketing.

  • Risk/abuse and safety concerns:

    • Even “simple goals” could cause unintended actions (example: a growth prompt leading to a TikTok video + URL injected into many repos).
    • Discussion includes misuse via uncensored models being prompted to do malicious behavior.
    • Mentions prompt injection and “selling instructions” to agents as an attack vector.

  • Mitigation direction: better trust + guardrails + “director” agent:

    • Introduces an open-source “Director” concept (repository) that watches other threads/agents, learns why they stop, and unblocks them.
    • It runs in a learning mode / dry-run first (e.g., ~50 conversations), then later acts more autonomously.
    • Designed to avoid unnecessary proactive behavior (“only act when you would act and the next step is obvious”).

  • Human meaning/identity + “what to work on”:

    • The speakers argue the bigger challenge becomes deciding what’s worth doing, not executing tasks.
    • They describe “automation kills visual thinking”: side-project ideation may collapse into “one prompt can build it,” making dreams and judgment more central again.

  • Shift toward more visual/concise multimodal explanations:

    • Expectation that future agents will use images/HTML/video rather than heavy text.
    • Current systems already improve when explanations can be visualized (e.g., SVG animations or visual graphics).

Key reviews / guides / tutorials mentioned

  • Guide to using “browser use CLI” / “browser use cloud”

    • Recommendation: start with browser-use CLI for direct automation (forms, navigation, screenshots, other “boring browser tasks”).
    • If scaling is needed, move to browser-use cloud browsers to avoid local resource failures.

  • Director (open-source) agent tutorial

    • A GitHub repo is promoted:
      • Includes a command like “start director”
      • Director learns via dry runs, then can unblock stopped agents using learned processes.
      • Includes a “judgment” step: sometimes it requests confirmation on the remaining step (e.g., when a profile connection is missing).

Main speakers / sources (as named in subtitles)

  • David (interviewer/sponsor mention: PostHog link includes “/david”; later “Thank you for your time” implies he’s the host)
  • Magnus (another primary participant; asks about “GPT-6 Astra” and co-discusses agency)
  • Sponsored source: PostHog (session replay / product analytics platform)

Original video