Will Weinbach on Autonomous Social Media Agents and Learning by Experiment

Social media strategist Will Weinbach joins Reid Hoffman and Parth Patil to discuss autonomous agents, content distribution and learning by experimentation. A preview includes an agent that posted four times while he slept and an experiment running an agent on a television.

The episode introduces a token grantee program offering selected creators $1,000 a week in tokens, a choice of tools and access to custom agents. It is intended to help creators close the gap between an idea and a finished work.

Patil says he met Weinbach on Clubhouse around five years earlier and was struck by how much energy he put into learning. He later added Weinbach to his Replit plan to see what a creative person without an engineering background might build.

Weinbach says he initially used ChatGPT for generic tasks such as writing an essay. Patil encouraged him to identify pain points in his own workflow and build an answer. He moved from prompting to automations and agents that perform tasks he would rather not do.

He compares his approach to learning music production: technical details can be valuable, but concentrating on them too early sometimes distracts from making anything. He says he does not understand much of the code and believes a beginner’s perspective helps him experiment.

Weinbach also admits that he is willing to take risks without knowing all the potential damage. Patil contrasts this with his own caution when OpenClaw became popular: he wanted to contain it on a separate Mac Mini before installing it.

Patil says Weinbach tried OpenClaw immediately and explored remote agents that can operate persistently. Around six weeks later, Weinbach introduced him to Hermes. Patil frames both as early, evolving systems and asks how Weinbach experienced the transition.

Weinbach says he first tried OpenClaw on a virtual private server on Patil’s advice, but did not know how to manage the server. He then installed it locally, accepting the risk to his computer. He was interested in tailoring the agent’s language and work style.

After finding Hermes through a recommendation on X, he tried it and preferred the visibility it gave him into agent activity, skills and changes. He says he values the ability to personalize its working style and retain context in files over time.

Hoffman asks how he discovers tools and what he expects next. Weinbach says he follows social media for user recommendations and tests many products; most do not become part of his workflow. He currently finds one agent simpler to manage, but imagines specialized agents for different parts of life and work.

He describes using social media knowledge and exploring graph-based retrieval to help agents work in unfamiliar domains. Patil characterizes this as building a knowledge graph before tackling a subject. They treat the approach as a developing technique, not a guarantee that an agent’s answers will be correct.

Patil says persistent memory and written skills can let an agent adapt without a new underlying model. In his description, the agent records relevant information and revises instructions as a user’s needs change; he regards the speed of this process as interesting.

Asked where autonomy paid off, Weinbach describes a social agent that selects and edits licensed video, adds text and copy, schedules posts and reads analytics. The goal is an iterative workflow in which performance data informs subsequent choices.

An early failure involved a company account. Weinbach says a poorly specified prompt caused the agent to publish about four posts back-to-back around 4 a.m. He was asleep, and was relieved when he saw that the content was on brand.

His lesson was to set clearer parameters in the initial loop, including what the agent should not do. The incident illustrates the difference between automating a workflow and reliably controlling its timing and behavior.

Hoffman asks what more autonomous content might mean for social media. Weinbach says audiences can be skeptical of AI-generated material and that the format matters. More broadly, he looks at why people watch, skip or share, and at how performance data can inform those decisions.

He walks through the workflow: a licensed library supplies thousands of images and videos, filtered by rules such as duration, visual style and the absence of an existing text overlay. Duplicate checks keep the system from repeatedly selecting the same material.

A copywriting skill analyzes the selected clip and proposes text based on guidance about what has worked. The workflow edits the video and prepares it for a calendar. The steps are separate skills assembled into one longer process, rather than a single command.

Weinbach says an API-enabled scheduling tool lets him publish across platforms. An Instagram/Meta integration gives him performance data such as average watch time and skip rate. The system uses feedback to test formats against a goal; these are his reported workflow and preferences, not independently verified performance results.

He says the workflow grew from individual frustrations: browsing the library, editing, writing copy and scheduling each took time. He uses Claude for this process and suggests the skill files might be portable to another compatible agent, subject to its setup.

For a stranger experiment, Weinbach set out to run Hermes on his television instead of buying a Mac Mini. Patil was skeptical and asked why he would put an agent on the TV instead of using one to control it. Weinbach says he got it running and sent Patil a message around 5 a.m.

Weinbach sees the experiment as a prompt to imagine agents on more devices, though the interview does not establish that every connected device can run one. Patil adds another example: an agent building a feature during a sales call could shorten the gap between selling and making an idea.

They discuss placing an agent in a group chat or Slack channel so multiple people can contribute context. Patil expects team agents to evolve with their collaborators, while presenting this as his view of an emerging way of working.

Hoffman asks what younger social strategists understand that older marketers might miss. Weinbach says generations grow up with different formats; his own experience is rooted in short-form video and an attention economy.

Weinbach argues for a broad hook that can reach people beyond an existing niche. He would rather have a post reach a million viewers and generate 20,000 profile visits than have a million followers but only 10,000 views per post. He presents this as a distribution preference, not a universal rule.

He suggests several reasons a person may share: expressing identity, connecting with someone, signaling knowledge or responding emotionally. A creator should think from the viewer’s perspective rather than only from the brand’s objective.

For a short video, Weinbach asks what will engage a viewer in the first three seconds, the next three, and the payoff at the end. His example is a nine-second video someone might want to watch again. Hoffman notes the focus is on seconds rather than minutes.

Weinbach says his own attention span has shortened and that he mainly watches longer content in the form of television episodes. He expects a similar trend more broadly; the interview presents this as his impression, not a measured finding about everyone.

Weinbach asks how companies might assess a young applicant whose ability to use AI tools is stronger than the experience on a résumé. Hoffman does not expect the résumé to disappear, but thinks its components and evidence of capability could change.

Hoffman imagines trusted AI systems that observe work and help validate particular abilities, alongside references from people who know the candidate. He envisions a more dynamic professional identity; this is a proposal about a possible future, not an established hiring system.

Weinbach asks how a college student could gain an advantage in a difficult job market, suggesting mass applications as one possibility. Hoffman argues for a more targeted demonstration: use AI to solve a problem specific to an organization and show what the applicant can contribute.

Hoffman gives his assessment of the causes of a difficult labor market and treats AI as an opportunity for young applicants. Those causal claims are his interpretation, not a conclusion established by this episode. His practical advice is to demonstrate a concrete capability rather than simply generate cover letters.

Patil says Weinbach’s familiarity with short-form social video lets him quickly spot problems such as text obscured by an app overlay or weak retention. He sees a parallel in early AI experience: practicing with new systems can give people examples to show even before job titles are established.

Patil recommends sharing what one has learned and building a visible body of work instead of blindly sending hundreds of automated applications. He believes a network can then see the person’s capabilities and initiate relevant conversations. Hoffman agrees, and the group closes the interview.