Agent Fleets
Multiple agents with separated responsibilities: one detects, one decides, one acts, one independently verifies the work actually happened. Verification is a separate agent with its own source of truth, not the actor grading itself.
What I build, and how it gets into your systems.
Four kinds of work, one method, and a public lab you can watch run. Start with the problem, because the fix follows from it.
Adoption was the easy part.
Whether you deployed something two years ago or you’re still deciding, the hard problems are the same four, and none of them are about model capability. They’re the operational questions a demo never has to answer, and they are engineering problems.
04 failure modes · all four are silent
These four problems have a public record. Read the receipts → Every claim on this site is checked at the source. Read the verification log →-
Nobody can tell you if it’s working
An agent that reports its own success is not a monitored system. Without independent verification against a source of truth, you learn about failures from customers, or from a number that has been wrong for a quarter before anyone checked it.
-
It degrades silently
Model updates change behavior. Upstream schemas drift. Prompt performance decays. None of that throws an exception. Output quality declines slowly, and unless something is measuring it, nobody notices.
-
It never got integrated
Demos run on a CSV. Production runs on your ERP, your ticketing system, your document store, and three things that only exist because someone built them in 2011. Connecting to those systems is most of the work.
-
Nobody owns it at 2 a.m.
Firms that build AI systems generally do not operate them. When the thing breaks eight months after handoff, the team that built it is on another engagement and your staff owns a system they didn’t design. Ask the succession question before signing: if the vendor vanished next quarter, could your own people run what is left?
What I actually build.
These are the four categories of work engagements fall into, not a product catalog. The assessment determines which ones matter for your operation, and usually it’s fewer than you’d expect.
01
Agent Systems
The hardest thing to buy off the shelf
Autonomous Operations
Agents with real authority to act on production systems: triage, remediate, escalate, document. Bounded by explicit policy, instrumented end to end, with human-in-the-loop where the blast radius warrants it.
Verification Architecture
The layer most implementations skip. Independent grading against ground truth, drift detection, and an audit trail that survives a compliance review.
Evaluation Harnesses
Regression suites for non-deterministic systems, so a model swap is a measured decision rather than a leap of faith.
02
Operational Automation
Where the recovered hours are
Document & Data Pipelines
Contracts, invoices, claims, intake forms, and field reports read, extracted, validated, and written into the systems of record, with a confidence threshold that routes exceptions to a human instead of guessing.
Back-Office Workflow
The multi-step, multi-system processes that quietly consume half an FTE: reconciliation, follow-up, routing, approvals, status chasing.
Systems Integration
Wired into what you already run: ERP, CRM, ticketing, telephony, document stores, and the internal tools nobody has documentation for.
Operations Dashboards
One screen showing what’s happening across the business in real time, including what the automation got wrong. Built to be acted on, not admired, whether it’s the dispatcher watching the board or the executive’s morning brief.
03
Custom Software
When the tool doesn’t exist yet
Internal Applications
Purpose-built tools for how your business actually works: field apps, customer portals, internal assistants, designed and coded from scratch, in your accounts, in your name.
Customer-Facing Products
Full products with payments, analytics, error tracking, and the operational plumbing that keeps them alive. I’ve shipped and run one myself.
Legacy Interfaces
The API that should exist but doesn’t. Getting modern systems to talk to the thing you can’t replace this year.
Security-Conscious Builds
Prompt injection defense, credential handling, tenant isolation, and audit logging treated as engineering requirements, not a checkbox added before launch.
04
Operate & Maintain
The part nobody else stays for
Production Monitoring
Systems watched around the clock with alerting on degradation, not just outage. Nothing I build fails silently. That’s the whole thesis.
Model Migration
Frontier models ship constantly. As better, cheaper, or safer ones arrive, I move your systems onto them against a regression suite, so what you bought improves instead of aging.
Incident Response
When it breaks at 2 a.m., that’s my department. Behind that is a decade of enterprise reliability engineering, at a scale most consultants have never operated at.
Accuracy Reporting
A monthly report of what the systems handled, what they escalated, and what they got wrong, with the receipts. The wrong-answer section is in every report, including the clean months.
AIistheeasypart.Gettingitintoyoursystemsisthejob.
The demo always works because it runs on a spreadsheet. Your operation runs on an ERP, a ticketing system, a document store, and the in-house tool nobody has documentation for. Connecting to that, safely and without a rip-and-replace, is most of what you’re actually paying for. Here’s how I approach it.
-
Meet systems where they are
A modern API or a nightly CSV drop: both are a way in
A clean API, an aging database, an SFTP folder of exports, a screen only a human has ever clicked. Each has a path in. Mapping your actual systems and finding the least invasive one, with the honest cost of each, is part of the assessment.
-
Read before write
It proves itself on your real data before it touches anything
Integrations start by reading from your systems and verifying the output against them. You watch it be right on your own data before it’s ever allowed to write back, held to the same verification standard the rest of my work is held to.
-
The API that should exist
Getting the new thing to talk to the one you can’t replace
When there’s no clean integration point, building the adapter is the work, and it’s work I’d rather do than pretend your legacy system isn’t there. The goal is to make what you already run more useful, not to sell you a migration you didn’t ask for.
-
In your accounts, in your name
You own the connection, not a box I rent back to you
Credentials are yours, the integration lives in your tenant, and when the engagement ends you own it outright. No proprietary middleware sitting between you and your own data.
The Walker Method
Enterprise-grade method. One senior engineer. No pyramid.
Every engagement follows the same five phases, the structure large firms use for AI transformation work, run by the person who will actually build and operate the system. The difference isn’t the framework. It’s that the senior engineer is present at every phase, including the two the big firms hand to someone else: the build and the operate.
-
01
Discover
Interviews and process mapping with the people doing the work.
-
02
Diagnose
Findings report: pains quantified, readiness scored 1 to 5, honest ROI.
-
03
Roadmap
A prioritized plan and fixed-price proposal, in writing, by a named day.
-
04
Build
Shipped against a written scope, in your accounts. You own it.
-
05
Operate
Monitored around the clock, reported monthly, mistakes included.
Phases 01–03 are the paid assessment. The report is yours whether or not we build anything together.
The lab is public.
Every pattern I sell, from verification fleets to autonomous ops, monitoring and full products, I built and run in production first, on my own infrastructure and my own dime. Most vendors show you a deck about what they’d build. Here you can watch mine run, and watch it catch its own failures.
Agent Fleet · Autonomous Ops
Incident Platform
A multi-agent system with the authority to detect problems, triage them, and restart production infrastructure on its own, then grade its own work against real metrics so it can’t lie about whether the fix held. The reference implementation for how I build agents with real authority.
Read the case study →Detector · Triage · Executor · Postmortem agents
Verification Architecture
The Red Button
A working, public demonstration of the three-pass Extract → Verify → Critique architecture. An agent fleet you can watch grade its own work, with a button that injects a silent failure so you can watch it get caught. Most vendors won’t show you how their systems fail. Here it’s the front door.
Press it yourself →Interactive fleet demo · Independent verification · Failure injection
Operations Dashboard · Fictional Sample
The Executive Morning Brief
The state of a company, meaning cash, receivables, service operations, and the three things that need the boss today, on one screen or texted to a phone at 6:30 AM. The sample runs on fictional data that regenerates daily, including the part that matters most: it flags what it couldn’t verify instead of guessing.
See today’s brief →Daily regeneration · Ranked priorities · Phone delivery
-
SITREP
Monitoring Agent
A hardened real-time monitoring dashboard aggregating threat and status feeds into one screen, with honest partial degradation when an upstream source goes dark instead of a blank panel. The pattern behind the operations dashboards I build.
Live feeds · Security-hardened · PWA
-
TariffIQ
Data Product
A live SaaS product with payments, analytics, error tracking, and an automated weekly newsletter. End-to-end proof I can ship and operate a customer-facing product, not just an internal integration.
Stripe · PostHog · Sentry · Beehiiv
-
NetWatch
Watchdog Agent
Always-on network monitoring that sweeps, captures, and reports on everything touching a network, quiet until something matters, which is the hard part.
Scheduled scans · Traffic capture · Alerts
-
Enterprise SRE
Active Practice
I currently hold a Staff SRE role keeping large-scale enterprise platforms online, which means the incident response, monitoring, and reliability patterns I build for clients aren’t things I did years ago. They’re things I did this week, at a scale most consultancies have never operated at. Clients get current production discipline, not a slide about it.
Staff-level · Rapid response · Production ops
What AI consulting costs in 2026
Other firms' published prices, checked at the source, in one table next to mine.
Mill test reports, read properly
The eight blocks, the ten tells, the limits my gate enforces, and where automated reading fails.
Field ticket to invoice
The seven places the money leaks, why portals reject, and what a recorded fleet caught and refused.