Jonata Santos
Complete AI Agent Development Workflow with Hermes and Research Loop
> how to build anything rn: > > - get a hetzner, do, or hostinger vps
Field guide / sub-domain
2
sources in this field
Updated June 13, 2026
Current thesis
The shortest path to orientation.
The editing surface itself is being rebuilt for agents. A new category of terminal emulators is emerging specifically for agentic workflows — vertical tabs, better notifications, and embedded/scriptable browsers optimized for long-running sessions — with libghostty becoming a platform for terminal innovation as the terminal layer is reinvented for the AI coding era. Decode pushes the same convergence inside Claude Code, embedding a browser and whiteboard for a visual feedback loop where developers annotate UX issues while the agent codes. UI exploration is being decoupled from git: UIFork lets developers fork UI variations within a single branch and flip between them in the running app, pointing at a new "design-in-code" category that bridges Figma-style exploration and production code, while Tool UI renders JSON tool outputs as inline, narrated, schema-stable surfaces inside chat (solving raw-text result dumps) and Nuxt UI templates reflect the broader pull toward starting from designed baselines rather than primitives. Cross-device development is maturing: Orca IDE on MacBook and phone with Tailscale enables seamless access to Hermes and other agent setups from any device. Even "headless" still needs human eyes — Astropad Workbench is a high-performance remote desktop for monitoring agents on headless Mac Minis from an iPad or iPhone.
Evidence board
Claims worth carrying forward01Deploy Hermes agent on Hetzner/DigitalOcean/Hostinger VPS with Codex auth configured for GPT-5.5 fast mode (no reasoning) for optimal performance
02Implement memory persistence using Gbrain or custom vault with QMD + SQL to store research findings and project learnings throughout development cycles
03Use `/grill-me` skill in Hermes to systematically uncover project unknowns before starting development, saving all results to memory vault for reference
04Install Orca IDE on MacBook and phone with Tailscale networking for seamless cross-device development environment
05Define clear evaluation criteria for every project to determine success, then iterate with Hermes until all evals pass while capturing learnings
06Claude Code can generate static HTML files for reports, explorations, code structure documentation, and mockups as primary outputs
07Install Claude Code skill by asking it to clone a specific repo, then request it to make any HTML page interactive to enable live editing
08Claude Code launches a localhost server that allows in-browser commenting on HTML pages, automatically updates content based on comments, and provides guided tours of changes
Adjacent fields
Key voices
Latest evidence
Jonata Santos
> how to build anything rn: > > - get a hetzner, do, or hostinger vps
Paras Chopra
> My favorite way of interacting with Claude Code is to have it generate static HTML files as outputs (reports, explorations, code structure, mockups etc.) > > I wanted to iterate on the file by commenting in browser and having Claude update the output live.