I build and govern fleets of AI agents that ship real systems.

A real governed agent run. An agent team scaffolds a CLI: every phase implemented, independently reviewed, then gated before commit.

Final gate: all five falsification criteria pass.

orchestrate: scaffolder build · tap to play · enable JavaScript to play

How I run agents

The orchestration model

Every run follows the same structure. A boss holds the mission and never writes. A manager owns the phase sequence and the state file. Workers are disposable; they receive a brief and report back. Adversarial reviewers are always separate from the writers they review. Human gates are hard stops, not suggestions.

Agent orchestration topology A directed graph showing the orchestration model. Boss delegates to Manager, who dispatches parallel Workers. Workers feed a Reviewer. The Reviewer reports to the Human Gate. The Human Gate approves Output. Boss Coordinates the mission Manager Runs the phase sequence Worker A Executes one phase Worker B Executes a parallel phase Reviewer Adversarial review, never the writer Human gate Hard stop, not a suggestion Output Shipped artefact

The map shows the structure used by the missions documented on this site.

Three Claude Code terminals running in parallel under one workflow
A live three-terminal workflow.

A run, documented

How a build actually goes

  1. The brief

    Gate cleared. Worker briefed.

    The wireframe gate passed. I issued the S2 brief to a disposable Sonnet worker: Astro skeleton, token system from the design plan, content collections wired, every page rendering real copy. The brief specified the Orchestra case page first, because that was the hardest screen. The worker had no prior context. It had the brief, the design plan, and the untouchable files.

  2. Delegation

    Worker builds. I stay out of the way.

    The worker built the full skeleton: layout, tokens, fonts, four routes, view transitions, the claims audit. It flagged two items it was uncertain about and kept going. That is correct behaviour. I read the output before touching anything.

  3. A gate fires

    Two defects. Both caught before publish.

    Manager verification found what the worker had partially flagged. Artifact links pointed at a GitHub remote that does not exist. A token and cost figure appeared in case copy, sourced from a roadmap file rather than a Ledger export. Both fail the falsification clause. Neither reached the published page.

  4. Recovery

    Amendments committed. Rescanned.

    Dead links removed. The figure withheld pending a real Ledger export. Em dashes found in source comments stripped. The amendments were committed, the build rerun, and every scan repeated from zero: dash scan, banned-word scan, locked file byte check. All clean.

  5. Review

    Mitch confirmed both calls.

    Both manager rulings went to the copy gate. Artifact links stay out until Orchestra has a real public remote. The figure reaches the page only through a real Ledger report embed. Both confirmed. The run closed the same day it opened.

The numbers

What the runs show

This site is one of them. The numbers below come from a ledger my own agents built, and not one of them was typed by hand.

4,609,452,561 Total tokens
195 Sessions
Input 7,603,607
Output 41,975,764
Cache read 4,249,920,585
Cache write 309,952,605
Product build collaboration
1,848,692,118 23
Client site A
740,553,907 4
Orchestra
725,985,344 60
Personal projects
349,561,810 2
Portfolio site
328,734,182 10
Small-model R&D
267,944,852 15
Knowledge base
207,247,118 16
Internal build sprints
47,786,866 2
TOMSSPYHQ
37,774,684 1
Design system
17,657,672 1
Workflow calculator
15,387,680 1
Internal tooling and eval runs
12,045,992 56
Misc shell sessions
10,080,336 4

Exported from real runs · 15 May 2026 to 13 Jun 2026

Selected work

What got built

Phase 2 underway · 2025-26

ORCHESTRA

A terminal HUD that watches the Claude Code CLI from out-of-band taps and puts every permission decision in the operator's hands.

Open the case
Tested prototype · 2026

SLM4SMB

A small-language-model appliance that reads service-business enquiry emails and books them on local hardware, with no cloud and no per-call fee.

Open the case
Live prototype · 2025-26

DIAGNOSTIC BUDDY

A mobile-first AI diagnostic assistant for automotive training that turns messy technician input into a structured diagnostic pathway.

Open the case
Working v1 · 2026

ORCHESTRATE

A CLI that scaffolds phase-gated, worktree-isolated, eval-first agent missions, built by an agent team running the discipline it encodes.

Open the case

More systems

Authentic PrizmaCore brain animation

Active

Second Brain

A Claude-native business vault with custom skills: working memory for research, workflow diagnosis, and reusable systems.

Shipped

Forgiveness Letter

A small, finished web app built in one unattended agent run. A personal one.

Open the app

Active

Pedal Builder

Design a guitar pedal in the browser and watch the circuit respond as you build. A collaboration build.

Shipped

TOMSSPYHQ

An offline arcade of classic and 3D browser games, built for a mate.

Active

DAY BUILD

A Claude Code skill library: seventeen product-building skills packaged as an installable marketplace.

Operating principles

How I work

No number exists unless a tool produced it. Every figure ships from a real export or it does not ship.

Every agent runs behind a human gate. Permission decisions stay explicit, logged, and mine.

Define failure before you build. Each system carries a written test for what would prove it broken.

Whoever builds it does not review it. Review is always a separate seat.

Incidents are evidence, not embarrassments. Every failure becomes a blameless postmortem and a new control.

State lives in files, not memory. Any run can crash and resume cold from what was written down.

Get in touch

Build with me

hello@prizmacore.com