The personal AI assistant just got affordable. And bendable. Here's the orchestration playbook I run three businesses on.

Two announcements this week looked unrelated. One: Hermes got lighter — plugin-driven, built to be bent to your operation. Two: an independent benchmark (25 real tasks) ranked it the most cost-efficient agent harness — $0.46 per completed task, 80% pass rate, second-fastest.

They're the two halves of one shift. I'm not a developer. I run three businesses — AI, tattooing, meal prep — on one Hermes system, as orchestrator and PM. When the assistant gets both affordable and bendable, what changes isn't the tool. It's who can afford to run one properly.

Here's the playbook I actually follow. It's not a product tour. It's how the orchestrator of a personal agent system operates — with the real surfaces from my setup.

1. The two announcements —

First, the product. The v0.11.0 "Interface release" is the clearest signal yet of the direction: plugins can now register slash commands, dispatch and intercept tools, rewrite tool results, transform terminal output, and add dashboard tabs. Live theming on every surface. Nous also published an official guide for building custom plugins.

Read that list again. It's not "we added features." It's "you can now bend this thing to your operation." Hermes is moving away from a one-size-fits-all tool and toward an assistant you assemble yourself.

Second, the economics. In Composio's 2026 benchmark of agent harnesses (25 real tasks), Hermes passed 20 of 25 — 80%, the third-highest pass rate. It was the second-fastest by median runtime (164 seconds). And it had the lowest cost per successful task of any harness tested: $0.46. Total estimated cost: $9.28. The report's own words: one of the strongest choices for value.

2. The connection nobody makes —

Lighter and cheaper look like separate updates. One is about how the tool feels. The other is about what it costs. But they answer the same question: can a normal operator afford a personal AI that fits their operation?

A year ago, the honest answer was no. Either the tool forced you into someone else's workflow, or it was too expensive to keep on. Now both barriers fell at once.

Form and economics are the same change. The plugin direction is how it looks. The compression and the $0.46-per-task number are what it costs. Together they say: the personal assistant is here, it's cheap enough to keep, and you can shape it to your work instead of shaping your work to it.

3. The orchestration playbook —

This is the part that matters. Affordability and bendability are the entry ticket. What you do with them is orchestration — and that's a discipline, not a feature. Here's how I run it.

3.1 Start with one role, not a team.

The instinct is to spawn a swarm. Wrong. Start with a single agent with one genuinely distinct job — research, drafting, coordination. Add agents only when a role is truly different from the ones you have. Parallelism helps only when subtasks are independent; otherwise it's coordination overhead disguised as progress.

My roster grew one role at a time, each with a real job:

3.2 Keep the hierarchy bounded.

You are the orchestrator; your agents are one level deep. Delegate with bounded depth — leaf agents that do their job and stop, never spawning an unbounded tree. Two levels, never more. A command chain stays checkable. A pyramid collapses under its own handoffs.

3.3 Let the work compound.

The value isn't any single call — it's that the same agent holds the same context across days. Give each role its own memory and its own pinned model. One agent, one job, continuity. That's what turns a tool into a colleague.

3.4 Route, don't babysit.

Orchestration is deciding which agent acts, in what order, with what shared context — then letting it run. My job is sequencing and routing, not re-prompting every step. When something's stuck, I escalate. I don't micromanage.

3.5 Gate only what's irreversible.

The gate pattern: the agent prepares and presents; I approve; it executes. Nothing publishes, sends, or pays without me. But I gate only the actions that are high-cost or irreversible. Gate everything and the gates become theatre — reviewers rubber-stamp and the boundary stops meaning anything.

3.6 Keep the record.

Every task, every handoff, every run has a trace — a durable board and a log. When a job goes wrong, the record is the impartial witness. I read it. I don't argue with it.

3.7 Price it before you build it.

The compression benchmark is why any of this is affordable — cost per completed task is now $0.46. So I price a task before I scale it, cap the expensive paths, and scale only the work that earns its keep. Affordability is a design decision, not a default.

3.8 Show the seams.

Keep the honest list of what never gets automated: anything that needs a body, a judgment, a relationship. The split between coordination and judgment isn't a limitation. It's the design that keeps the system real, accountable, and keepable.

4. The honest limit —

The assistant doesn't run my businesses. The work that needs a body or a judgment is still human — cooking, tattooing, deciding, selling. What the system holds is the thinking layer: the coordination, the context, the record, the drafts.

That split is the design, not a compromise. The new economics make the thinking layer cheap enough to keep. They don't remove the human.

5. The move —

The personal assistant is now affordable and bendable. The only variable left is whether you shape it to your operation — or let it stay a demo.

Cheap enough to keep. Bendable enough to fit. Orchestrated like a real team.

The one move this week: pick one role, give it its own memory, run it for a week. Keep the coordination. Keep the gates. Show the seams.