AI development · Agent integration · Workflow implementation
ControlStackAI is the engineering practice of Matthew Mangano. I design, build, and operate AI systems that hold real context, take real actions inside real tools, and leave a record you can review afterward — for teams that need the work finished, not another pilot.
Proof of work
This is not a concept for something I hope to build. It is my own working environment: one spoken request becomes retrieved context, a plan, delegated execution across specialist agents, and a consequential action that stops for human approval and leaves a receipt. The same architecture is what I build for clients.
Voice-first
The request is spoken. No prompt-chaining, no copy-paste between tabs.
Persistent memory
Decisions and constraints survive the session, so the system stops re-asking what it was already told.
Verified delegation
Specialist agents do the narrow work and report back with evidence, into one continuous thread.
Review before action
Anything consequential freezes, asks for approval on the exact action, and writes a receipt after.
Project case study · Agent Harness
I built a provider-independent runtime for agent work where the executor never grades its own homework. It runs against a definition of done, survives interruption, verifies the result separately, and leaves a tamper-evident receipt.
Explore the architecture01
Define the contract
Intent, checks, constraints, budget.
02
Execute durably
Local, API, or delegated agent.
03
Verify separately
Checks plus adversarial review.
04
Keep the proof
Signed verdict and evidence.
Services
Different entry points, same underlying job: get a model doing useful work inside systems that already have owners, constraints, and consequences.
01 — Build
Products with a model inside, built like software rather than like a prototype that got promoted. Retrieval that returns the right thing, structured output you can parse, evaluations that catch regressions, and cost and latency treated as requirements.
02 — Integrate
Coding agents and assistants are only as useful as their reach. I connect them to the repositories, tools, and services your team actually uses, with permissions that are scoped deliberately instead of granted by default.
03 — Automate
Take a process a person runs by hand every week and make it a supervised one. Map the real workflow — including the exceptions nobody documented — automate the load-bearing parts, and put a human at the decisions that deserve one.
Method
The same four beats the film walks through are how the engagement itself runs. AI work fails in predictable places: nobody wrote down what the system is allowed to assume, and nobody defined what "working" means before building it.
01
Watch the process as it is actually performed, not as the documentation describes it. Name the constraints, the data boundaries, and the steps that carry consequences.
02
One workflow, end to end, in production, early. A narrow system people use beats a broad one they evaluate.
03
Scoped permissions, approval on consequential actions, and an evaluation suite that fails loudly. Autonomy gets extended once it is earned, not assumed on day one.
04
Documented, tested, and runnable by your team. You should be able to change it without calling me — and call me because you want to, not because you are stuck.
Stack posture
The frontier moves every few months. Anything built tightly around one vendor's assumptions becomes a migration project the moment the pricing or the capability changes. I build the boundary first — your workflows and evaluations on one side, the model and harness on the other — so swapping either is a decision rather than a rewrite.
In practice that means working across the current generation of coding agents and agent runtimes, commercial and local inference, and the tool protocols that connect them — and being honest with you about which parts of that are stable enough to depend on.
Who you are hiring
I am an aerospace engineer. Before AI systems, my work was aircraft 6-DOF simulation, flight dynamics, control law development, and the verification that has to hold up when something flies. That background is the reason I build AI the way I do: define what correct means, instrument it, and never confuse a convincing output for a validated one.
I also run the thing in the film. My own infrastructure is a fleet of specialist agents on a declaratively managed Linux host, with persistent memory, scheduled autonomous work, and approval gates on anything that leaves the machine. Every failure mode I would otherwise discover on your project, I have already hit on mine.
You work with me directly. No account manager, no handoff to a junior team after the pitch.
Background
Engagements
Most people should start with the audit. It is fixed-fee, it ends with something useful whether or not we work together again, and it means neither of us is guessing at scope when the build starts.
Start here
Fixed feeOne to two weeks
I map how the work is done now, find where a model realistically helps and where it does not, and rank the opportunities by value against difficulty.
You get: a workflow map, a ranked opportunity list with honest effort estimates, a recommended architecture, and a build plan you can execute with me or without me.
Build
ScopedFour to eight weeks
One workflow taken all the way into production, with the evaluation suite, the permissions, and the approval gates that make it safe to leave running.
You get: a working system in your infrastructure, source and tests in your repositories, runbook documentation, and a handover session with the people who will operate it.
Operate
MonthlyRetained, rolling
For teams running agentic systems in production who want someone accountable when the model, the pricing, or the workflow changes underneath them.
You get: continuous extension and maintenance, evaluation and cost monitoring, model migrations handled, and an engineer who already knows the system.
Pricing is quoted per engagement after a short scoping conversation — there is no useful honest number before knowing what the work is. Tell me the workflow and I will tell you whether it is worth automating at all.
Contact
The most useful first message is a concrete one: what the process is, who runs it today, how often, and what breaks when it goes wrong. I reply personally, usually within a business day.
Prefer email? [email protected]