Platform

The infrastructure behind long-running agent work

Work that crosses many sessions needs more than a model. It needs durable structure, controlled execution, reliable interpretation, and composable inference.

Four boundaries

ldgr — structure and history

Keeps work, runs, evidence, validation, and decisions outside the model so a project can survive context resets and restarts.

ldgr.run

agentctl — execution and output

Supervises agent processes, keeps full command logs, and returns bounded summaries instead of pouring raw output back into the next context.

github.com/hydra-dynamix/agentctl

Grist — interpretation

Parses documents, repositories, source code, structured data, and model output into typed forms that downstream systems can inspect.

github.com/hydra-dynamix/grist

Erasmus 3 — inference and workflows

Composes model calls, tools, speech, translation, and transformation services into typed multi-step workflows. It is private infrastructure, not a public service.

Erasmus 3

Erasmus pipeline interface showing a Codex agent, text-to-speech, speech-to-text, and model-output parser in one workflow — A typed pipeline combines an agent with speech and parsing services. The interface exposes the order, input type, output type, and run state of each step.

A stack, not one bundled product

These tools have different jobs. They do not all write to one ledger, and they do not need to be installed as one package.

We describe an integration only where it exists. The shared principle is narrower: model output is candidate state, execution is bounded, and durable records stay available after the model context ends.

What the stack can produce

Session Board is the complete build example: a multi-service event platform produced by a long-running Pi agent loop structured with ldgr. Follow the build from brief to working system.