Underpass AI

Memory, coordination, and execution infrastructure for reliable AI agents.

Underpass AI helps agents recover context, coordinate specialist work, execute through governed tools, and leave auditable evidence behind.

We do not build foundation models. We build the operational substrate around them: navigable memory, event-driven coordination, governed execution, evidence, policy, and observability.

Explore GitHub Read the KMP article Contact

Problem

Agents are getting better at work, but weak infrastructure makes them hard to trust.

Modern agents can read code, call tools, run tests, and operate across complex workflows. But most agent systems still lose context between runs, repeat failed attempts, hide state inside frameworks, and produce weak audit trails.

  • They lose relevant context between executions.
  • They repeat failed attempts because past process memory is not navigable.
  • They cannot always explain what evidence supported a decision.
  • They execute tools without enough governance or inspection.
  • They are difficult to observe, debug, and audit in production-like systems.

Platform model

Three infrastructure planes around the model.

Underpass separates memory, coordination, and execution so each part can be inspected, tested, and evolved independently.

Domain event
Specialist agents
Memory Underpass KMP
Coordination Underpass Choreographer
Execution Underpass Runtime
Evidence + memory update

Agents should not start every task from zero. They should recover scoped memory, understand previous attempts, coordinate the next step, execute through controlled runtimes, and write evidence back into the system.

  1. Domain event
  2. Specialist agents
  3. Underpass KMP restores scoped memory
  4. Choreographer coordinates deliberation and work
  5. Runtime governs tool execution
  6. Evidence is recorded
  7. Memory improves for the next event

Components

Public infrastructure components.

KMP and Runtime are the main public technical assets. Choreographer is public and in active development.

  1. Public memory plane

    Underpass KMP

    Kernel Memory Protocol for temporal, multidimensional, auditable AI agent memory.

    • Scoped memory with about and dimensions.
    • Temporal traversal: goto, near, rewind, forward, trace, inspect.
    • Evidence-backed deterministic retrieval.
    • Explicit relations and provenance.
    • Typed gRPC API plus MCP adapter over the same semantics.
    • Kubernetes/Helm deployment path.
    • Adapter-based persistence for graph, key-value, and event roles.
    View Underpass KMP on GitHub
  2. Public execution plane

    Underpass Runtime

    Governed execution plane for tool-driven AI agents.

    • Isolated workspaces.
    • Governed tool execution.
    • Policy checks before execution.
    • Telemetry and evidence trails.
    • Adaptive tool recommendations.
    • mTLS- and Kubernetes-oriented deployment.
    • OpenAPI/gRPC-oriented runtime surface.
    View Underpass Runtime on GitHub
  3. Public coordination plane

    Underpass Choreographer

    Deliberation engine for specialist agent councils.

    • Councils propose, critique, revise, validate, and score.
    • Declarative YAML ceremonies: states, steps, roles, guarded transitions.
    • Optional LLM-as-judge ranks proposals by intrinsic quality.
    • The debate is replayable as an OpenTelemetry trace.
    • Deliberation metrics: judge discrimination, winner scores, token cost.
    • Typed gRPC + AsyncAPI contracts, MCP adapter, Kubernetes/Helm.
    View Underpass Choreographer on GitHub

Why now

Agentic systems are moving from demos to operational workflows.

LLMs are becoming capable enough to use tools, inspect code, run commands, and participate in real engineering workflows. That creates a new infrastructure problem: memory, governance, auditability, and observability need to become first-class parts of the system.

  • Software engineering is one of the first domains where agentic workflows can produce measurable feedback.
  • Tool-using agents need runtime isolation and policy checks.
  • Long-running agents need memory that can be navigated, not just retrieved.
  • Human operators need evidence trails, traces, and failure classification.
  • GPU-backed local and distributed inference is a strategic technical direction.

Writing

From the lab notebook.

Longform notes on the ideas behind Underpass AI: navigable agent memory, auditable multi-agent deliberation, and the economics of running agents in production.

  1. AI infra

    I let my GPU workers shut themselves down after five minutes

    Running evidence-bound agent reviews on ephemeral AWS GPU workers while preserving traces, logs, metrics, and decision artifacts after shutdown.

    Underpass AI · English
  2. Choreographer

    No queremos agentes que contesten. Queremos decisiones que se puedan auditar

    A multi-agent design meeting as a declarative ceremony: proposals, peer critique, an LLM judge — and the whole decision replayable as an OpenTelemetry trace in Grafana.

    Dev.to · Spanish
  3. Operator

    Operator: cuando responder no basta

    Training Qwen 0.5B with LoRA to emit exact Kernel Memory Protocol actions from visible memory state, without turning it into a general reasoning model.

    Dev.to · Spanish
  4. KMP

    Building Kernel Memory Protocol: navigable memory for AI agents

    Why agent memory has to be temporal, multidimensional and auditable — and what a kernel-style protocol for it actually looks like in practice.

    Dev.to · English
  5. Event pipeline

    What an event-driven agent pipeline looks like when you trace it end-to-end

    Walking a full agent pipeline as a stream of events: who emits what, where decisions get coordinated, and what the trail actually proves about behaviour.

    Dev.to · English
  6. Economics

    Why event-driven agents reduce scope, cost and decision dispersion

    The architectural and economic case for event-driven coordination over chatty agent loops: smaller scope, cheaper retries, fewer dispersed decisions.

    Dev.to · English
All posts on dev.to/tirsogarcia

Open source

Built in the open.

Underpass AI is developed around open infrastructure principles. The public GitHub organization exposes the technical direction, the core repositories, and the foundation for future community collaboration.

Memory

Underpass KMP

rehydration-kernel · Rust

View Underpass KMP repo →
Execution

Underpass Runtime

underpass-runtime · Go

View Underpass Runtime repo →
Coordination

Underpass Choreographer

underpass-choreographer · Rust · Apache-2.0

View Underpass Choreographer repo →
Explore the GitHub organization

Founder

Founder-led technical infrastructure.

Underpass AI is created by Tirso García Ibáñez, a software architect with 15+ years of experience across software engineering, architecture, distributed systems, cloud-native platforms, and technical leadership.

The current focus is AI infrastructure for agentic systems: navigable memory, governed execution, Kubernetes-native runtime foundations, and production-oriented architecture patterns for LLM agents.

linkedin.com/in/tirsogarcia

Contact

Talk to Underpass AI.

Interested in reliable agent infrastructure, technical validation, open-source collaboration, or early-stage product conversations?