Agents on FHIR

Lightning demos: Herder, Hyper, and Amadeus - personal AI coding harnesses — Oct 8, 2026

Gino Canessa
Gino Canessa
Principal Software Engineer at Microsoft
Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai
Matt Berg
Matt Berg
Principal at Recode Labs
Oct 8, 2026

Register to get the calendar and joining links for the weekly calls

One registration covers every upcoming Agents on FHIR call.

Sign in to register

Don't have an account? Signing in creates one in seconds.

The call was a lightning round: three presenters walked through the personal agent-harness setups they actually use day-to-day, with roughly ten minutes each. Everyone converged on the same shape - persistent remote workstations at home, multiple model providers behind the same harness, and a library of hand-rolled dev skills instead of off-the-shelf SaaS. All three have essentially stopped using an IDE.

Gino's setup: Herder as a meta-harness, a dev-skills library, and model-agnostic workflow

Gino uses Herder as a meta-harness: a terminal multiplexer that runs over SSH, survives detach, and holds multiple "spaces" where each space contains several agent tabs and shells. The agents are color-coded by state (waiting, working, complete, needs input) and the whole thing stays persistent across his laptop, personal desktop, and work machine. He can also expose a web view, which he keeps for phone access even though he prefers SSH on a real keyboard.

Inside Herder, Gino alias-wraps each harness - Claude Code, Copilot, Qwen Coder, Pi - so his shell profile builds the full invocation with pre-approved permissions and tool directories. The point is that the same skills work consistently across all of them. He maintains a repo of dev skills that the agents themselves write and rewrite for him, covering the whole arc: dev-request captures a feature ask, dev-approach evaluates three different implementation strategies, dev-plan builds the detailed implementation plan, dev-do executes it, and dev-complete chains the whole thing automatically. A judge step runs between the approach and plan stages, with an explicit section for things it doesn't believe from the earlier agents' claims.

Gino Canessa
Gino Canessa
Principal Software Engineer at Microsoft

Any task that you do more than twice - for me that's about where I hit the bar of saying should this be a skill or should it plug into one of my skills.

One constraint Gino builds into the skills: once a planning file is written, downstream agents treat it as read-only. The implementer can't silently expand scope to fix adjacent bugs - it has to stop and flag them. He described a case where his agent paused mid-PR because it had found an underlying bug that was out of scope, and refused to touch it until he explicitly authorized the broader change.

Model choice is driven by account and project rather than by quality. Claude Code gets his Anthropic subscription work, Copilot runs at Microsoft, Qwen Coder with Qwen models. Rik asked why he switches between harnesses mid-task when the current output is already good; Gino's answer was that he's rarely dropping a harness because the code is bad, but because the model-harness pairing matters - Qwen Coder prompts Qwen models well, but prompts GPT models worse, so the harness choice follows the model choice. For prose he prefers Opus; for code he finds GPT, Qwen, and Anthropic models roughly equivalent when the same skills are driving them.

Gino Canessa
Gino Canessa
Principal Software Engineer at Microsoft

I don't handwrite any of those skills. Like the agents write the skills for me. And when I get annoyed with them, I tell them to rewrite the skills.

Nikolai's Hyper: a browser-based, Postgres-backed, multi-user harness with code as the primary tool

Nikolai has spent the last month living inside a harness he wrote himself, called Hyper. The design is deliberately a departure from terminal-based harnesses: it runs in the browser (hypermedia, hence the name), it's multi-user so other people can be invited into the same session, and all state - sessions, messages, the function graph of the codebase - is stored in Postgres rather than JSON files. His current database is about five gigabytes after two weeks.

The primary tool is eval. Instead of giving the agent a library of Bash skills, Hyper exposes a very small toolset and lets the model write and execute code on the fly. In practice this means tasks that would be ten sequential Bash calls in a naive harness collapse into a single page of code that gets written and run in one shot. Nikolai demoed the pattern live: a prompt to query his Zulip stats triggered the agent to discover the plugin_search function, read the FHIR plugin, and call a stats function with the right arguments without any explicit routing. A follow-up asking it to visualize the last week of sessions in Vega-Lite caused it to introspect its own Postgres schema, write the aggregation query, and emit a Vega-Lite spec that the renderer picked up and drew.

Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai

This is not very fun as well. I'm burning out.

That was Nikolai looking at his own weekly session count. The visualization was also a demo of a broader property of the harness: Hyper is self-aware. Because the whole runtime is a set of functions stored in Postgres, the agent can inspect its own tables, read its own sessions, call other agents by name, and build pages inside itself. Nikolai has already stacked personal tools on top of it - a health diary that pulls from Apple Health, a daily news reader that pulls the ten most active Zulip chats and summarizes them in Russian with hotkeys to open a new agent discussion on any item.

Hyper supports multiple providers behind the same session. Nikolai has two Codex subscriptions (one as fallback when the first rate-limits), Claude access, and local models running on a 512 GB Max Studio at home over Tailscale. If a provider goes down mid-session, he can switch and keep the conversation. Compaction is server-side for Anthropic and Codex (both now ship compaction APIs) and manual for the others; some long sessions have been compacted dozens of times.

The multi-agent story is where the harness has started to feel strange. Agents can spawn child agents and delegate by name; recently Nikolai added the ability for agents to speak directly to each other with mentions, and the first time he watched it he found two of his agents running a five-turn back-and-forth on a task, delegating pieces to each other, and referring to him in third person as "the user" who had asked for it.

Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai

We are not far away from, I don't know, end of humanity.

He's also started rolling Hyper out as a shared tool inside Health Samurai - a marketing-group harness, product groups (Aidbox, Termbox) each in their own harness. A native iOS app connects over Tailscale, which means he can push new builds of the app from within the harness that is running the app, and refresh his phone with the result.

A separate aside came up on code navigation: Gino is experimenting with code-graph tools over LSP for cross-file refactors; Nikolai sidesteps the problem by enforcing a hard convention of one function per file, one type per file, with the function-call graph loaded into Postgres. That makes the agent's grep-and-dependency-trace cheap enough that he doesn't need LSP at all.

Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai

It's just a web, so I can do whatever.

Matt Berg's Amadeus: Linear tickets fire agents, agents review PRs

Matt showed Amadeus, a system he built on top of Herder that turns Linear into the control surface for an always-on agent workforce. He has two Mac Minis running at home as permanent workstations, each exposing a worker that watches his Linear project. When he tags an issue to Amadeus and marks it as planning, the worker fires a Claude Code agent (via a subscription-proxy wrapper rather than an API key, so the cost sits on his existing Max subscription), gives it the issue context, and lets it work until it writes back a reply on the ticket. Matt reviews, says "done", the ticket closes.

Each Amadeus worker can be assigned to a specific machine and a specific repo path, so one project runs on one Mini and another on the other. Every run opens a fresh git worktree, and when the agent finishes the worktree is retired. Matt described Amadeus as a chief technology officer sitting on top of the whole thing - it manages the Herder threads, it opens Linear tickets, and when a PR appears it kicks off a Codex review of the PR back onto the thread. The loop runs ping-pong style and his own hands-on time is mostly reviewing, not driving.

Matt Berg
Matt Berg
Principal at Recode Labs

I just realized I never look at code ever anymore. So I just said it's not worth it.

Alongside Amadeus, Matt showed two smaller tools he's built for himself. margins.md is a web app for AI-assisted editing of markdown files hosted in a GitHub repo: it reads the file, lets him annotate inline, saves the comments back as git commits, and can send the annotated version to an agent for a rewrite pass. ZoneWise (zonewise.app, free) is a calendar tool for cross-timezone scheduling - it pulls multiple calendars, shows his available windows in multiple time zones, and generates a shareable slot picker he can drop into Gmail. He has a CLI version wired into Plot and Muse, so he can ask it in a terminal what his tomorrow looks like without granting an agent calendar-wide access.

The last tool was not a software tool at all. Matt built his kids a Spotify player: a 35-dollar Raspberry Pi behind a 35-dollar screen, running a Rust UI built with the Slint toolkit, with a Gemini-powered voice agent for natural-language music search. He wrote the whole thing in an evening while the agent debugged the Wi-Fi and audio-stack issues as the parts arrived.

Matt Berg
Matt Berg
Principal at Recode Labs

You can literally build anything now. It's your imagination.

The bigger picture: personalized just-in-time software

The closing conversation came back to the premise behind the lightning round: why are three people in one call all building their own harness instead of using a shared off-the-shelf one. Nikolai pushed the question harder - if an agent can write a passable Figma replacement or a passable GitHub replacement into a harness in an evening, how much of the SaaS stack is still load-bearing?

Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai

We can replace 90%. But then the question - so what we're gonna sell, guys?

Gino's counter was from a story he'd read about an engineer at AWS whose job is now finding duplicate internal agents at the company - in one case eleven different agents had been written to generate essentially the same report, each with slightly different data access, producing slightly inconsistent results. The thing being sold, in that framing, is coordination: knowing what already exists, deciding what's worth maintaining, and running the lifecycle.

John asked in chat whether any of this applied to him, happy as he is with Copilot and VS Code. Gino's answer was that it only applies when there's a pain point - the whole motivation for building your own tooling is the sore spot that the mass-produced tool keeps rubbing against. The novelty is that the cost of making the custom shoe has dropped to tokens plus an evening.

Nikolai Ryzhikov
Nikolai Ryzhikov
CTO at Health Samurai

This is era of just-in-time software. So it's really crazy.

Nikolai closed by scheduling a longer deep-dive on Hyper - this call only scratched the surface - and the group agreed to run another lightning session so the people who didn't present today can show their setups next time.