ep 06 field notes
Show Us Your Agent Skills / EP 06 / guest dossier
SKYLAR PAYNE WICKED DATA PALMER HERMES REVIEW QUEUE

SKYLAR PAYNE

Skylar's personal assistant is named Palmer: a Hermes agent on a Mac Mini on his desk that runs a local tech community of about 80 people, answers his email, and curates wedding-planning notes he has never touched. Then success created a problem: 30 half-baked runs coming back with garbage to review. So he built hermes-workflows, where the agent writes Python and the human is one of the primitives. ask() puts Skylar's decision inside the program the same way agent() calls a model, every answer lands as structured data, and one recurring job has already been distilled into an open-weight model of his own.

▶ WATCH SEGMENT $ PERSONAL-AGENT-OPS $ DYNAMIC-WORKFLOWS
80 PEOPLEin the tech community Palmer runs, weekly events included 30 THINGSrunning and "coming back with garbage I need to review" MAC MINIPalmer lives on Skylar's desk, always on, reachable through a channel ZERO EDITSPalmer curates the Obsidian wedding notes; Skylar has never touched them ASKthe human answers as a typed step, same contract as an agent OWN MODELthe Obsidian updater got distilled into an open-weight model
EP 06 · SKYLAR PAYNE · Palmer, always-on operations, and workflows with human checkpoints, live on stream

"I'm gonna call them slop repos because I largely have not reviewed the code. Mostly I just make sure it works for my purposes."

The open source repos he demoed shipped anyway. They work for his purposes; the code review can wait. 00:48:40

HIS HERMES AGENT RUNS THE OPERATIONS

the personal-agent-operations workflow: always on, and named Palmer

Palmer is a Hermes agent on a Mac Mini on Skylar's desk, always on and reachable through a channel rather than a terminal he has to open. After moving back to a hometown without a strong tech community, Skylar put Palmer in charge of one: event setup, notifications, email replies, hackathon coordination. Now there are weekly events, "and it's all mostly just managed. I don't have to think about it."

Palmer publishes into artifactd, an HTML workspace Skylar built because he thinks visually. It stores the documents the agent creates, with tags, names, and search. Palmer made the episode intro there, audio included, with an ElevenLabs voice it picked for itself.

Obsidian holds the operational memory. Palmer curates wedding-planning notes spanning vendors, contracts, tasks, people, and links back to Gmail, drawing on Skylar's CRM. "It somehow pulled out who the best man is and who the maid of honor is without me ever saying anything about that."

Skylar Payne showing Palmer's artifact workspace with generated documents
Palmer's artifact workspace gives generated work a durable place to live, with tags, search, and protected cards. [00:59:58]

"I often realize, okay, I have 30 things running coming back with garbage I need to review."

Agents made half-baked starts cheap, and vague tasks multiply past human review bandwidth. He answered with typed steps and human checkpoints. 00:46:31

THE HUMAN IS A FUNCTION CALL

the hermes-dynamic-workflows workflow and the hermes-workflows-creating skill: the agent writes the Python, and ask() is you
Notice what agents skipWhen multiple agents shared a repo, they sometimes skipped the worktree or separate-checkout step, no matter what the instructions said. 01:03:48
Write the workflow as code"The sort of agentic interface here is really just code." The Hermes agent writes Python with agent steps in it. 01:05:55
Compose with the primitivesagent runs a subagent with a name, prompt, and return type; parallel and pipeline arrange the work; a loop is something that does and something that checks. 01:06:17
Type the human in with askask requests the same structured output from a person that an agent would return. 01:07:28
Live in the review queueApprove the markdown-rendered plan, pick one content angle from several, or request changes: "no, not that. Do X instead." 01:08:50
Run long projects on triggersNo continuously running process: trip planning resumes when a trigger or human decision lands, which keeps Skylar on schedule too. 01:11:01
Skylar Payne showing a Hermes workflow run DAG for a content creation workflow
A workflow run as a DAG: the content creation workflow with completed steps for visual aids, content artifacts, and workflow completion. [01:08:15]
Skylar Payne showing the Hermes Workflows dashboard with a review queue and saved artifacts
The Hermes Workflows dashboard: runs, artifacts, and the review queue where human decisions resume the workflow. [01:08:50]

"One of the problems is that prompts and skills are effectively suggestions."

A skill can work in a clean context and fail in a crowded one: "the performance of a certain skill or prompt isn't constant. It depends on what else is in the context." 01:03:48

EVERY APPROVAL IS TRAINING DATA

each click in the review queue gets saved, and the traces are already replacing models

Every workflow step has a structured output, and the review queue layers human feedback onto it. That turns daily work into data: "that now allows me to easily sample traces for outline drafting." The traces feed prompt optimization and distillation into smaller, cheaper, faster models.

The updater that monitors his email, SMS, and calendar to maintain Obsidian projects and people records has been distilled into an open-weight model. Skylar checks the open model scene every three or four months; local models earn their place when the task repeats.

For coding workflows, the structured output always includes a VS Code SSH link so he can inspect the code, though "I don't do that often unless something seems suspicious or surprising to me."

Skylar Payne showing Obsidian notes curated by Palmer for wedding planning
Palmer-curated Obsidian notes turn personal operations into a project surface: people, decisions, vendor threads, and linked follow-up work. [01:01:59]

"Hermes is constantly looking at the trajectories and saying, these things are kind of related. This should be a skill."

The harness writes its own skills by watching what Skylar does, then culls the similar and unused ones over time. 00:53:56

WHAT HE SHIPPED

two workflows, a published skill, two open repos, and a post
workflow

personal-agent-operations

An always-on assistant with real duties: community events, email, artifacts, and memory, reachable through a channel.

workflow

hermes-dynamic-workflows

Agents write Python against structured primitives, humans answer typed questions from a review queue, and traces get saved.

skill

hermes-workflows-creating

Skylar's published authoring skill: how an agent should write a workflow against the primitives.

repo

artifactd

The HTML workspace runtime Palmer publishes into: tags, names, search, and generated documents, including the episode intro.

repo

hermes-workflows

The open source plugin behind the demo: workflow execution, DAG views, artifacts, review queue, and receipts.

"The first few weeks of OpenClaw was a beautiful high that I've never found again."

Then releases got brittle. OpenClaw could rewrite its own config into an invalid state and go dark: "you have to go send Codex or a rescue team in to fix it." 00:52:34

SKYLAR'S STACK

what keeps Palmer running
Hermesthe always-on harness Palmer runs on Obsidianthe memory surface Palmer curates Codexlocal coding, and maybe Hermes's replacement someday OpenClawthe earlier harness, beautiful high included ElevenLabsPalmer picked its own voice for the intro DSPyone inspiration behind the workflow plugin artifactdthe HTML workspace Palmer publishes to hermes-workflowsprimitives, DAGs, and the review queue

"Making it easy for humans is very similar to making it easy for agents."

When an agent returns an overcomplicated plan for a small change, the code was usually already tangled for people. Give both the curl request, the endpoint, and the expected response. 01:23:17