AI Boost

The Boost Levels

I built the AI frameworks and workflows that took my delivery from 1–3x earlier to 10–100x recently — in production, at scale, with high quality.

* If you are still struggling the single chat with any AI, think about shift to Context Engineering.

* If you never thought about 100x, read all of my stories already happened in production as scale.

Chat with one AI10–50%

Context engineering, grown from prompt engineering1–3x

Agent engineering: harness, fleet, loop10–100x

1x3x10x30x100x

My own delivery, stage by stage, on a log scale. My experience in my own work, not a benchmark.

How to Scale with Quality?

I build agent workflows, harnesses and skills in Claude Code daily, and I've run multiple agent-driven campaigns in production at scale on the Scan-and-Action Framework (SAAF). Everything gets reviewed before anyone else sees it.

Framework: Long-Horizon

I built a harness of my own, Long-Horizon, for agent work that runs for hours or days and survives interruption. It pairs a stateless dispatcher with a durable checklist, so the work outlives any single session.

In one proving run, it completed 32 checklist items across 21 sessions in 9 hours with zero human interventions — recovering on its own from a quota exhaustion, a two-hour operator absence, an abandoned 46 GB transfer, and two network outages.

How I run an agent team

Conductor: plans and assigns, counts the votes and never casts one · an hourly beat resumes the run after a failure

1Build11 designs, built and previewed locally
2Vote6 reviewers, 2 of them cold; each ballot a file
3Red team2 passes, separate from the builders
4Fix & re-checkby a script first shown to catch planted errors
5I picknothing ships before

One conductor plans and assigns; worker sessions build; the conductor counts the votes and never casts one.
It runs unattended overnight, with an hourly beat that picks the run back up after a failure.
Eleven designs were built and previewed locally, then chosen by a vote of six reviewers — two of them brought in cold, with no prior context — each ballot written as a file.
An independent red team, separate from the team that built them, argued how each design could hurt me, in two passes, before anything reached me.
Every fix was applied and re-checked by a script that had first been shown to catch planted errors.
Nothing ships until I pick.

Why the harness holds quality at scale

A charioteer driving two horses in one harness: the black horse is Claude Code with the local codebase context, the white horse is Rovo with the organizational context; the reins are prompts and control, the bridle and blinders are scoping and context, the yoke keeps the two aligned, and the traces carry the deliverables. — Image: an AI-generated illustration from my LinkedIn post.

Ride one AI horse on its own and the ride can get wild. In a harness, two horses pull one load in the same direction. The reins are my prompts and control, the blinders keep each horse on its own context, and the yoke keeps them in step. Claude Code brings the codebase, Rovo brings the organization's knowledge, and in one afternoon they took a design to what I'd call near perfect: Rovo's final review found every design decision traced back to the code or the docs. Read how →

The Flawless Design Pattern: Swiss-cheese model

Every check has holes, so I stack independent ones, a doer, a validator and a gatekeeper, and a bad change has to get past all three.

The Swiss-cheese model: a doer, a validator and a gatekeeper as three independent layers; a bad change passes the first two and is stopped at the gatekeeper. — Layers that each have holes, stacked so the holes don't line up.

Every result is reviewed by someone other than its author before it reaches a person — a second session, or an independent reviewer with no stake in it.
I seat an adversarial reviewer for anything that matters, whose job is to find how it fails.
I don't trust a check until it's been shown able to fail first, on a planted fault.
The work outlives any single session: checkpoints to git, a durable checklist, and a dispatcher that resumes after a crash, a quota limit, or an outage.
Claims get pulled from the source, not from memory, in the same turn they're made — a relayed fact stays marked as relayed until it's confirmed.
Anything that touches a live system runs with a second session watching each change, with a rollback written first.

Read the case study →

As an AI coach, I advocate an AI mindset shift and share how to boost 100x

I lead the BrainTrust knowledge-sharing series — about six years now, from around 2020. It's run to 150+ sessions, 61 of which I've delivered personally, 40+ hours in all. I'm Principal of API University, with more than 30 internal learners, and I'm ramping a 15-person team now.

channel