Artificial intelligence program

HEFESTO

Multi-Agent Harness Engineering

Engineering Reliable Multi-Agent Systems through Harnesses, Runtimes, Governance and Control Planes

$600

10 phases · 15 modules · capstoneStudent repo · not bridle2 recorded tutorials / month · spoken in Spanish, English subtitles

Course roadmap · 10 phases

Phase 1

HARNESS

Model → agent → harness

Phase 2

CONTEXT

Context, state, tools

Phase 3

ROLES

Role engineering

Phase 4

MULTI-AGENT

Orchestration

Phase 5

VERIFY

Verification + failure

Phase 6

MULTI-HARNESS

Adapters, common governance

Phase 7

CONTROL PLANE

Design, not the Kratos product

Phase 8

GOVERN

Security + observability

Phase 9

COST

Tokens and equal-compute

Phase 10

CAPSTONE

Governed multi-harness system

Ten technical phases. Fifteen academic modules (I–XV) live inside them. Coursework clock is not sealed. Same path as the accordion below.

Mission and vision

Mission

Train engineers to design the system that controls the agents: harnesses, runtimes, governance, and a control plane on paper — governable, verifiable, observable, reproducible.

Vision

A Governed Multi-Agent Harness System that separates ROLE, MODEL, HARNESS, PROVIDER, TOOLS, STATE, POLICY, and EVIDENCE, and that does not add agents by default.

Why HEFESTO

HEFESTO studies the engineering required to turn one or more AI models into governable, verifiable, observable, and reproducible agentic systems. It does not teach prompt writing, and it does not stop at creating an agent. The object of study is the system that controls the agents.

Holding the model fixed and changing the harness can change performance materially. The result still depends on Model × Harness, not on either one in isolation.

Who it is for

AI engineers, agentic-systems engineers, software and platform architects, ML / MLOps / LLMOps engineers, coding-agent developers, and technical leads who must govern multiple models and providers. Prior multi-agent experience is not required. Python or Go, Git, APIs/CLIs, basic LLMs, and tool calling are.

What it is not

Not Kratos core. Module IX designs a minimal control plane on paper. Seal, seats, receipt, and `/start` stay in Kratos / bridle. Labs are your repo.

What you'll be able to do

Design a Governed Multi-Agent Harness System that does not collapse ROLE, MODEL, HARNESS, PROVIDER, TOOLS, STATE, POLICY, EVIDENCE
Decide Single Agent vs Multi-Agent vs Multi-Harness without assuming more agents are better
Run a harness-swap experiment with model and task held constant
Separate verification from production, and refuse correlated-failure folklore
Design a minimal control plane on paper — without teaching the Kratos product

Course content

Phase 1 — HARNESS 2 modules

Module 1 — From Model to Agent System Outline + lab L1

Model ≠ Agent · AGENT SYSTEM = MODEL + HARNESS + ENVIRONMENT Read

Lab L1 — Harness Swap Experiment Open lab

Module 2 — Harness Anatomy Outline + lab L2

Observation, context, loop, tools, memory, state, verification, execution Read

Lab L2 — Minimal Agent Harness Open lab

Phase 2 — CONTEXT 2 modules

Module 1 — Context Engineering for Agents Outline

Context as a system resource · Context Fidelity Test Read

Module 2 — State, Checkpoints & Recovery Outline

Session resume ≠ checkpoint resume ≠ reconstruction ≠ hot migration Read

Phase 3 — ROLES 1 module

Module 1 — Role Engineering Outline + lab L5

ROLE ≠ MODEL ≠ HARNESS ≠ PROVIDER Read

Lab L5 — Role–Capability Matrix Open lab

Phase 4 — MULTI-AGENT 2 modules

Module 1 — Single-Agent vs Multi-Agent Outline

Agent Necessity Test · when another agent is justified Read

Module 2 — Multi-Agent Orchestration Outline

Topologies, handoffs, anti-patterns Read

Phase 5 — VERIFY 2 modules

Module 1 — Verification Engineering Outline

Producer ≠ judge · Correlated Failure Test Read

Module 2 — Failure Engineering Outline + lab L14

Failure taxonomy · DETECT / CONTAIN / RECOVER / AUDIT Read

Lab L14 — Failure Injection Campaign Open lab

Phase 6 — MULTI-HARNESS 1 module

Module 1 — Multi-Harness Engineering Outline + lab L8

Agents → different harnesses → common governance Read

Lab L8 — Harness Adapter Interface Open lab

Phase 7 — CONTROL PLANE 1 module

Module 1 — Agent Control Plane Outline + lab L9

Execution plane vs control plane — design, not Kratos Read

Lab L9 — Minimal Agent Control Plane Open lab

Phase 8 — GOVERN 2 modules

Module 1 — Security & Governance Outline

Capability Gate — no capability because the model can use it Read

Module 2 — Observability & Evidence Outline

Trajectory + lineage + decision evidence Read

Phase 9 — COST 1 module

Module 1 — Cost & Token Engineering Outline

Equal-Compute Benchmark · SAS vs MAS Read

Phase 10 — CAPSTONE 1 module

Module 1 — Capstone Outline

Governed Multi-Agent / Multi-Harness System Read

Course content is Phase → Module → Chapter using HEFESTO's ten-phase path. Module numbers restart inside each phase. Labs remain chapters inside their module. Lecture bodies and lab procedures are not authored yet.

Instructor

Elmer Niño

Founder, VTAlgo Group · Systems Engineer · MSc in Computer Science

Algorithmic trader since 2009, running trading as an enterprise. HEFESTO is harness engineering: the system that controls the agents — not another agent, and not the Kratos product.

Programa de inteligencia artificial

HEFESTO

Multi-Agent Harness Engineering

Ingeniería de sistemas multi-agente fiables mediante harnesses, runtimes, gobernanza y planos de control

$600

10 fases · 15 módulos · capstoneRepo del alumno · no bridle2 tutoriales grabados / mes · hablados en español, subtítulos en inglés

Mapa del curso · 10 fases

Fase 1

HARNESS

Modelo → agente → harness

Fase 2

CONTEXT

Contexto, estado, tools

Fase 3

ROLES

Ingeniería de roles

Fase 4

MULTI-AGENT

Orquestación

Fase 5

VERIFY

Verificación + fallos

Fase 6

MULTI-HARNESS

Adapters, gobernanza común

Fase 7

CONTROL PLANE

Diseño, no el producto Kratos

Fase 8

GOVERN

Seguridad + observabilidad

Fase 9

COST

Tokens y equal-compute

Fase 10

CAPSTONE

Sistema multi-harness gobernado

Diez fases técnicas. Quince módulos académicos (I–XV) viven dentro de ellas. El reloj de cursada no está sellado. El mismo recorrido que el acordeón de abajo.

Misión y visión

Misión

Formar ingenieros para diseñar el sistema que controla a los agentes: harnesses, runtimes, gobernanza y un plano de control en papel — gobernable, verificable, observable, reproducible.

Visión

Un Governed Multi-Agent Harness System que separa ROLE, MODEL, HARNESS, PROVIDER, TOOLS, STATE, POLICY y EVIDENCE, y que no añade agentes por defecto.

Por qué HEFESTO

HEFESTO estudia la ingeniería necesaria para convertir uno o varios modelos de IA en sistemas agentivos gobernables, verificables, observables y reproducibles. No enseña a escribir prompts, y no se detiene en crear un agente. El objeto de estudio es el sistema que controla a los agentes.

Mantener el modelo fijo y cambiar el harness puede modificar materialmente el rendimiento. El resultado sigue dependiendo de Model × Harness, no de uno de los dos aisladamente.

A quién va dirigido

AI engineers, ingenieros de sistemas agentivos, arquitectos de software y de plataforma, ingenieros ML / MLOps / LLMOps, desarrolladores de coding agents y technical leads que deban gobernar múltiples modelos y proveedores. No se exige experiencia previa en sistemas multi-agent. Sí Python o Go, Git, APIs/CLI, LLMs básicos y tool calling.

Qué no es

No es Kratos core. El Módulo IX diseña un plano de control mínimo en papel. Sello, asientos, receipt y `/start` quedan en Kratos / bridle. Los labs son tu repo.

Lo que vas a poder hacer

Diseñar un Governed Multi-Agent Harness System que no colapse ROLE, MODEL, HARNESS, PROVIDER, TOOLS, STATE, POLICY, EVIDENCE
Decidir Single Agent vs Multi-Agent vs Multi-Harness sin asumir que más agentes es mejor
Correr un experimento de harness-swap con modelo y tarea fijos
Separar verificación de producción, y rechazar el folklore de fallos independientes
Diseñar un plano de control mínimo en papel — sin enseñar el producto Kratos

Contenido del curso

Phase 1 — HARNESS 2 modules

Module 1 — From Model to Agent System Outline + lab L1

Model ≠ Agent · AGENT SYSTEM = MODEL + HARNESS + ENVIRONMENT Read

Lab L1 — Harness Swap Experiment Open lab

Module 2 — Harness Anatomy Outline + lab L2

Observation, context, loop, tools, memory, state, verification, execution Read

Lab L2 — Minimal Agent Harness Open lab

Phase 2 — CONTEXT 2 modules

Module 1 — Context Engineering for Agents Outline

Context as a system resource · Context Fidelity Test Read

Module 2 — State, Checkpoints & Recovery Outline

Session resume ≠ checkpoint resume ≠ reconstruction ≠ hot migration Read

Phase 3 — ROLES 1 module

Module 1 — Role Engineering Outline + lab L5

ROLE ≠ MODEL ≠ HARNESS ≠ PROVIDER Read

Lab L5 — Role–Capability Matrix Open lab

Phase 4 — MULTI-AGENT 2 modules

Module 1 — Single-Agent vs Multi-Agent Outline

Agent Necessity Test · when another agent is justified Read

Module 2 — Multi-Agent Orchestration Outline

Topologies, handoffs, anti-patterns Read

Phase 5 — VERIFY 2 modules

Module 1 — Verification Engineering Outline

Producer ≠ judge · Correlated Failure Test Read

Module 2 — Failure Engineering Outline + lab L14

Failure taxonomy · DETECT / CONTAIN / RECOVER / AUDIT Read

Lab L14 — Failure Injection Campaign Open lab

Phase 6 — MULTI-HARNESS 1 module

Module 1 — Multi-Harness Engineering Outline + lab L8

Agents → different harnesses → common governance Read

Lab L8 — Harness Adapter Interface Open lab

Phase 7 — CONTROL PLANE 1 module

Module 1 — Agent Control Plane Outline + lab L9

Execution plane vs control plane — design, not Kratos Read

Lab L9 — Minimal Agent Control Plane Open lab

Phase 8 — GOVERN 2 modules

Module 1 — Security & Governance Outline

Capability Gate — no capability because the model can use it Read

Module 2 — Observability & Evidence Outline

Trajectory + lineage + decision evidence Read

Phase 9 — COST 1 module

Module 1 — Cost & Token Engineering Outline

Equal-Compute Benchmark · SAS vs MAS Read

Phase 10 — CAPSTONE 1 module

Module 1 — Capstone Outline

Governed Multi-Agent / Multi-Harness System Read

Instructor

Elmer Niño

Fundador, VTAlgo Group · Ingeniero de Sistemas · MSc en Ciencias de la Computación

Trader algorítmico desde 2009, que lleva el trading como una empresa. HEFESTO es ingeniería de harness: el sistema que controla a los agentes — no otro agente, y no el producto Kratos.