Concept

Reference agency

The reference organization shipped with ALDO AI. Use it as the example for every platform feature.

This page is the public-docs version of the agency README. The agency/ directory in the repository contains the reference set of agent specifications shipped with ALDO AI. Every file under direction/, delivery/, support/, and meta/ conforms to the aldo-ai/agent.v1 schema.

These are the agents the platform itself uses to build, review, and operate products — they are also the canonical examples authors reference when writing new agents for their own tenants.

#Layout

agency/
  direction/      Principal, strategy, architecture, program management
  delivery/       Engineers who produce artefacts (code, infra, models, pipelines)
  support/        Reviewers, auditors, writers, ops partners
  meta/           Agents about agents (author, evaluator, HR, historian)

Each agent YAML has a companion system prompt under <team>/prompts/<name>.system.md. Prompts stay with their agent so that promoting a new version always moves the spec and the prompt together.

#Conventions

LLM-agnostic. Specs declare model_policy.capability_class (reasoning-large / reasoning-medium / reasoning-small / local-reasoning / fast-draft); the router resolves a concrete provider at call time. No agent names a vendor model.
Privacy first. Anything that reads source or secrets is sensitive and carries a local-reasoning fallback so a tenant can keep work on-prem.
Reporting lines. principal reports to no one; product-strategist, architect, and program-manager report to principal; tech-lead reports to architect; delivery engineers report to tech-lead; support agents report to the function they partner with.
Eval gates are real. Every agent names at least one regression suite with a threshold the registry enforces at promotion time.
Spawning is narrow. Only supervisors spawn, and only the roles listed under spawn.allowed. Reviewers and auditors never spawn.

#Promotion

version: 0.1.0 across the board — these are the initial drafts. The registry refuses to promote an agent to a higher version without the referenced eval suites passing their declared minimums.

#Use it as a test case

Every platform feature should answer "does this help ALDO TECH LABS ship?". The agency is the reference organization — if a feature doesn't help these 26 agents do their job, the platform doesn't need it.