The Agent Network, Part 3 of 3: Revisit the Problem, Reuse the Answer

Keep the problem; reconsider the solution

A product idea can stall because the technology is inadequate while the underlying problem remains worth solving. This final part of The Agent Network conversation opens with an approach to that impasse: attach conviction to the problem, while remaining willing to abandon a particular solution. Revisit unsolved problems as AI changes what is feasible, with enough humility to recognize which ones exceed your capacity.

Keeping up with those changes takes a concrete form for one participant: roughly two hours of late-night YouTube, often playing while they code. Their schedule includes a bedtime around 2 a.m., seven to seven and a half hours of sleep, and an effort to avoid morning meetings. Watching talks suggests a task they want an agent to handle: extract the slides into a reusable deck, or distill the talk into another format. They believe the technology is ready; the slide-extraction agent remains a proposal.

Make a simple operation accessible

Agent.ai offers video transcript retrieval as an action primitive—a reusable operation an agent can invoke. Fetching a transcript may seem mundane to an engineer, but exposing it removes a practical barrier for someone who does not know how to retrieve it programmatically. Timestamps then connect relevant passages to footage that can be used in clips.

The participant describing this capability built their original agent for their wife, who wanted to combine several clips into a nonprofit-related compilation to send to a friend without using video-editing software. The task illustrates why a small operation matters: it can make a desired workflow accessible to someone who otherwise lacks the tools to complete it.

Another participant recommends Gemini for video understanding, citing its recently added YouTube support, and also mentions Meta’s video-understanding models. The recommendation reflects ease of use at the time of the conversation, rather than a measured comparison.

Image editing: the donut has to fit

An experimental Gemini Flash image model stands out to one participant as the most exciting AI development of the preceding week to ten days because it can edit images. Their test captures both the appeal and the remaining limitation: asked to replace an element of the HubSpot logo with a donut, the model made the substitution but sized the donut incorrectly. A correct substitution can still produce an unusable composition. They regard the capability as an early version; the exact model version is unspecified.

The same participant describes frustration with Stable Diffusion: trying a similar prompt produces a different image when the task calls for a controlled edit. They characterize editing as nearly impossible but acknowledge limited knowledge of the algorithms. This is an intuition about regeneration versus controlled editing, not a general technical limit on diffusion-based systems.

The discussion considers two directions for image tools: - Generation methods crossing domains: One participant contrasts diffusion’s association with images and autoregression’s association with language and next-token generation. They describe David Holz as betting on text diffusion for Midjourney’s path forward—an account of his position, rather than a direct statement from him. - Interfaces for editing and generation: Another praises Playground’s image editing and Photoshop-like generation interface, while explicitly saying they do not know whether its underlying approach is autoregressive.

For one participant, autoregressive image generation has reopened a field they had come to see as mostly pursuing larger models and greater detail. They view the new capability as a potential competitive threat to Photoshop and Canva. That is a prediction about the direction of image tools, rather than an established competitive outcome.

Protect time, then make answers reusable

The conversation turns from technical capability to personal capacity. The author of the post referred to in the closing exchange as Sorry Must Pass explains that it grew out of an overwhelmed schedule and guilt about declining requests—even a fifteen-minute startup-idea review introduced through a friend. Every acceptance consumes time available for something else. A refusal reflects that capacity limit, rather than a judgment on the requester or project.

Their default is to pass, following Derek Sivers’s hell-yes-or-no principle: reserve limited capacity for requests that earn an enthusiastic yes. They call the explanatory post their most personally useful piece of writing and still send it personally. At the time of the conversation, they automate neither email responses nor social posts.

Another participant suggests a complementary practice: publish answers to recurring questions. When people repeatedly ask for the same advice, a written explanation provides something useful to share without repeating the conversation.

That idea extends into dharmesh.ai, a personal knowledge agent built on Agent.ai. Its builder describes it as available and text-only at the time, with no audio or video. Written interactions supply its knowledge through a private email address.

The builder favors asynchronous exchanges over one-on-one meetings, while still joining team Zooms. They may spend twenty minutes writing a thoughtful answer when they can control the schedule, reserving the right to reuse it without attribution to the requester in the model or a public blog post. The email backlog becomes material for future answers, retrieved through a vector store. They report that this works surprisingly well for their own use, without providing a quantitative evaluation or explaining the retrieval implementation.

Keep audience value at the center

In the closing exchange, one participant says their approach to AI engineering was explicitly inspired by HubSpot’s inbound marketing: recognize a trend and build a substantial industry category around it. They then ask how an AI engineering community could go wrong.

The answer defines inbound marketing as using limited resources to create something useful for the intended audience. It contrasts the organizing principle with its failure mode: - Create audience value: Publish material that target customers or community members find useful, rather than interrupting their day to push a purchase. - Drift toward extraction: Lose that focus as the community grows and begin trying to squeeze more from the audience.

Making money is compatible with the mission, provided the core remains raising awareness and creating value for members. The respondent’s belief that markets eventually reward generosity is a personal philosophy, not a promised business result.

The recording closes with the address for the post about declining requests: sorrymustpass.org.