Point of contact
The applied-AI lab of Carlos Guzman · Live on Google Cloud
ACE SystemAgentic CognitiveEnterprise.My three-yearapplied-AI research program.
I built ACE to study how governed AI agents learn, specialize, and operate as one system. Steward is the top-tier agent that coordinates the rest.
Models used in the research
Claude
ChatGPT
Gemini
ACE research systemGoogle Cloud · Live
RESEARCH FILM 01
ORCHESTRATION · GOVERNANCE · CONTROL
ORCHESTRATION · GOVERNANCE · CONTROL
ACE-PRIME · RESEARCH PROTOTYPESteward governs
the agent system.LIVE EXPERIMENT
the agent system.LIVE EXPERIMENT
Research stackFrontier models inside one governed architecture.
Anthropic · Language modelClaude
OpenAI · Language modelChatGPT
Google · Language modelGemini
Approval before sensitive action
Access model
Access when a role ends
A working lab · Not a concept reel
Design.
Build.
Govern.
Operate.End to end.
LIVE
Google ADK · Gemini · Cloud Run
Google ADK · Gemini · Cloud Run
ACE-Prime · Top-tier research agent
Meet Steward.
Steward is the ACE reference implementation for governed orchestration: one top-tier agent that holds system context, coordinates specialist agents, and routes authority back to the human owner.
ACE-PRIME · TOP-TIER AGENT · HUMAN GOVERNEDSteward
The top tier of the ACE hierarchy
Steward demonstrates how one governed orchestrator can coordinate a wider system of specialist agents.
Can one agent coordinate many specialists?
The experiment studies context sharing, delegation, escalation, and performance across an agent hierarchy.
Observe, decide, coordinate
Steward interprets enterprise signals, assigns specialist work, and returns governed actions to the system.
Governance remains outside the model
Roles, access, escalation, and final authority remain explicit, reviewable, and controlled by the human owner.
ACE RESEARCH NOTE · STEWARD
“The model is not the system. The system is the governed relationship between agents, context, tools, and human authority.”
How Steward works
Describe the work.
Steward creates the role.
No prompt engineering. No model selection. No workflow diagram. The owner describes what is not getting done.
01
ACE stage 1 · Job definition
The pain point becomes a job.
Steward writes the job description from a description of the real operational problem.
OWNER
“Nobody's answering the enquiries that come in overnight, and by the morning half of them have gone cold.”
02
ACE stage 2 · Agent creation
The position is composed.
Steward defines the role, tool loadout, company context, reasoning thresholds, and the line it cannot cross.
STEWARD
“You need a Receptionist. Reads the inbox overnight, drafts replies in your voice, books into your calendar. Nothing goes out without your yes.”
03
ACE stage 3 · Onboarding
The owner says yes.
The specialist begins already oriented to the company. Steward passes down only the knowledge the role requires.
OWNER
“Do it.”
ACE specialist layer
Steward coordinates.
The agent system responds.
This research tests whether specialist agents can operate as governed roles inside one shared cognitive architecture—not as disconnected chatbots.
SupportCustomer SupportACTIVE ROLE
GrowthBD AnalystACTIVE ROLE
AnalyticsData AnalystACTIVE ROLE
OperationsWorkforce AnalystACTIVE ROLE
RevenueProposal WriterACCESS REMOVED
SupportCustomer SupportACTIVE ROLE
GrowthBD AnalystACTIVE ROLE
AnalyticsData AnalystACTIVE ROLE
OperationsWorkforce AnalystACTIVE ROLE
RevenueProposal WriterACCESS REMOVED
They work as one team.
Steward routes work across specialists while the owner maintains one relationship.
Nobody starts from zero.
Culture, policies, people, and voice belong to the company. Each role receives only what it needs.
No agent commands another.
The team is flat on purpose. Every line of authority returns to the owner.
ONE.
Shared context · System formation
Every agent begins with orientation.
This part of the research explores how explicit owner context can guide agent decisions without surveillance or uncontrolled inference.
Steward’s executive briefing
01
How the owner decides
Priorities, risk tolerance, and standards.
02
How the owner communicates
Tone, vocabulary, and working style.
03
What the owner expects
What proceeds, escalates, or always waits.
04
Leadership context
Provided directly, never inferred through surveillance.
Every specialist’s company orientation
01
Company and culture
What the organization does and the standards it upholds.
02
People and reporting lines
Who owns, approves, and receives escalation.
03
Policies and procedures
The approved sources required by the job.
04
Role, responsibility, access
A job description and only the systems it requires.
Need-to-know by design.
Steward holds the complete orientation. A receptionist may know the booking policy and escalation contacts. It does not need the financial plan or board materials.
booking policyescalation contactsfinancial planboard materials
Specialist layer · Research role catalog
Not just a model.
A governed position.
Each specialist is defined by competence, tools, authority boundaries, and evaluation criteria inside the ACE architecture.
FRONT OFFICE
ReceptionistCustomer SupportExecutive AssistantCommunications
REVENUE
BD AnalystSales DevelopmentAccount ManagerProposal Writer
MARKETING
Content MarketerDemand & TrendsMarket ScoutLead Generation
ANALYTICS
Data AnalystWorkforce UtilizationCoverage OptimizerInteraction Analyst
KNOWLEDGE
Research AnalystResearch AdvisorSubject Matter ExpertThreat Analyst
Agentic governance · The ACE lifecycle
Defined for a role.
Governed through every stage.
Purpose, access, evaluation, escalation, reporting, and controlled exit remain explicit—with governance enforced in code.
Drafts the email.
It writes into the owner's real drafts folder, in the owner's voice. A human presses send.
Does not spend money.
If a role needs a budget, the owner sets the ceiling. The role cannot raise it.
Does not rewrite systems of record.
Analysts read. They do not write. A role cannot change the numbers it reports on.
Does not give orders to other agents.
Agents cooperate; they do not command one another. Every line of authority returns to the owner.
How agents think
Six layers.
One cognitive architecture.
Explicit contracts, quality targets, and governance controls move each ACE agent from a prompt to a governed operational role.
Architecture in motion · Governed by designFrom human authority
to coordinated action.
The operating model becomes visible: authority, orchestration, context, reasoning, tools, and delivery working as one system.to coordinated action.
Infrastructure
Compute, storage, networking, auto-scaling, secret management, and observability.
Tools
Read, write, compute, communicate — registered, versioned, rate-limited, and logged.
Context
Knowledge bases, RAG pipelines, memory, and enterprise data for grounded decisions.
Reasoning
LLM cognition, confidence scoring, and escalation logic.
Orchestration
Sequential, parallel, conditional, and event-driven coordination.
Interface
APIs, chat, voice, dashboards, and webhooks — where humans meet the workforce.
Research hypothesis · Evidence before autonomy
Agents grow.
Evidence earns autonomy.
Every agent starts supervised, advances through measured performance, and works inside the same continuously governed operating system.
Cognitive career ladder
A structured path from task execution to governed strategic advice. No agent skips a tier.
ACE-CoreTask WorkerSUPERVISED
ACE-XSpecialistSEMI-AUTONOMOUS
ACE-PrimeOrchestratorAUTONOMOUS
ACE-TitanStrategic AdvisorFULL · GOVERNED
Observe. Decide. Execute.
ACE Engine interprets enterprise signals, resolves priority, and activates the next governed action.
Scheduling variance detectedLIVE
Compliance drift approachingROUTED
Attrition risk spikeESCALATED
OBSERVEDECIDEEXECUTE
Proof · Selected product portfolio
Concept to production.
Across real systems.
Agentic platforms, copilots, and governed workflows built across enterprise operations, infrastructure, family connection, legal technology, and education.
ACE System
The Agentic Cognitive Enterprise platform with hub-and-spoke orchestration, enterprise intelligence, voice, and operational dashboards.
acesystem.ai · Production
XmartGrid
Predictive asset intelligence for energy and water infrastructure, with an AI copilot and live digital twin.
xmartgrid.com · Production
AeroXmart
Seat-program intelligence and orchestration across lifecycle reviews, route structure, and configuration.
aeroxmart.com · In development
Vita
Daily multilingual voice connection with a concise family briefing and context for the next call.
vita.lucernas.ai · In development
GF Engine
Colombian child-support legal intake with a sequential agent pipeline and real-time compliance validation.
guzmanfossigroup.io · Pipeline topology
MySafeBud.AI
A K–12 AI tutor with parent-configured guardrails and a safety pipeline before a child sees a response.
mysafebud.ai · Safety architecture
ReversaMiEdad
A Spanish-language biological-age advisor with a hybrid agent architecture, running on Google Cloud.
reversamiedad.com · Production
MyBattleBuddy
Veteran-first, voice-first AI companion with medication support, a consented Battle Buddy Pact, a VA/DoD-aligned Safety Plan, and tiered human escalation.
mybattlebuddy.ai · Private beta
The Lab · Carlos Guzman
This ismy lab.
ACE System is where I take ideas from concept to production — designed, built, governed, and operated end to end by me. Everything on this page is real and running: a multi-agent workforce with governance enforced in code, deployed on Google Cloud, live right now at steward.acesystem.ai. Not a product. Not for sale. A working record of applied AI.
— Carlos Guzman
Carlos GuzmanENTERPRISE AI EXECUTIVE · BUILDER · OPERATOR