AI agents at AR Labs: the agents we run and what each does

AR Labs runs its own work with AI agents: language models run in a loop with tools, so they can act as well as reply. A few hold standing roles, with their own instructions and records, and keep working between conversations; many more are started for one task and end with it. This page lists them, what each does, how a person works with it and how far each has got, with a link to the page that describes it.

The agents

Agent	What it does	How a person works with it	Where it runs	State
Strategy agent	Runs the lab's planning reviews, yearly to daily; records decisions and passes direction down	Starts a review in a coding session and decides at each checkpoint	Inside a coding session on a laptop	Running
Marketing manager agent, Night Shift	Researches, recommends what to write and why, prepares writing briefs, coordinates and measures	Chats with it, sends requests through the shared mailbox, approves anything published, sent or spent	On a server, always on	Running, supervised; fuller structure designed
Vox Prime	The lab's identity in the shared messaging channel, so a laptop session can ask a resident agent for something and read the answer	A session posts and reads under it	Inside a coding session on a laptop	Running for sending and reading; automatic delivery designed
Coding agents	Write specs, tests and code, and review each change	Approves specs, answers at gates, merges	A fresh coding session for each step, on a laptop	Running
Research agents	Search the web and report with sources	Asks the question, reads the report	Started from a coding session	Running
Voice agents	Hold spoken conversations: a sales discovery call, a discussion of a document	Talks to them in a browser	A voice server reached through a browser	Salesbot running and public; VOX Presenter building

How they relate

A person sits at the centre and makes every consequential decision. The strategy agent, Vox Prime and the coding and research agents all run inside sessions on the laptop; the marketing manager agent runs on a server. Two channels connect the two places. The shared mailbox carries durable handoffs as files, and the messaging channel carries conversation, with a mention when one party asks another to act. The engineering of those channels is described on resident AI agents: a multi-agent system with roles.

flowchart LR
  P["A person"]
  subgraph LAP["Laptop: coding sessions"]
    SA["Strategy agent"]
    VP["Vox Prime"]
    CA["Coding agents"]
    RA["Research agents"]
  end
  subgraph SRV["Server, always on"]
    NS["Marketing manager agent, Night Shift"]
  end
  MB[("Shared mailbox")]
  CH["Messaging channel"]
  P <-->|"reviews, decisions"| SA
  P -->|"approves, merges"| CA
  P <-->|"conversation, approvals"| CH
  SA -->|"briefs"| MB
  MB <-->|"requests, replies"| NS
  VP <-->|"posts, reads"| CH
  CH <-->|"mentions, replies"| NS
  CA -.->|"starts"| RA

What makes an agent here

We call something an agent when it has four things.

• A standing role. A written statement of what it is responsible for and what it is judged on.
• Its own instructions and memory. Files it reads at the start of every run, and records it keeps, so its work survives the end of a session.
• Authority limits. What it may decide alone, what it must recommend and leave to a person, and whose records it may not change.
• A person approving anything outward-facing. No resident agent publishes, sends or spends on its own, and no coding agent merges its own work. A voice agent such as Salesbot talks to the public by design, inside a conversation whose stages and limits are fixed in code.

A script follows fixed steps and makes no judgement, so we call it a script even when it starts agents: the coding pipeline's driver is one. Between the two sit task agents, started for one step of one task. They have a role and instructions but, by design, no memory of their own between tasks: each step starts a fresh session that reads the task's records, which keeps one step's reasoning from leaking into the next. The resident agents described on this page add memory that persists and a channel through which others can reach them. How roles, runtimes and models fit together in a coding agent's harness is described in the anatomy of an agent harness.

Other agents the lab uses

In the coding pipeline

Every code change takes the same route, the coding pipeline, and each step is done by an agent with one role.

• The engineer agent builds a task from its written specification, test-first, in a worktree of its own. See test-driven development with AI coding agents.
• The test writer writes or extends tests for existing code or for a specification's acceptance criteria, in dedicated test passes.
• The architect works on system design and writes the designs and specifications other agents build from.
• The reviewer is a model from a second vendor, reached through OpenAI's Codex CLI, which reviews each change in a fresh context from the diff, the specification and the proof alone. See code review gates.
• The driver is a script, not an agent: it reads the task tree, starts a fresh agent session for each step, and stops at each gate that needs a person. See multi-agent orchestration, unattended agent runs and Coding Planner.

Research

• The research fleet splits a question across three, nine or 24 researcher agents of three kinds (web search with an academic emphasis, general synthesis, and a second provider's citation-grounded search) and combines their findings into one sourced report. See multi-agent research, and orchestrate for how a session decides whether to start agents at all.
• The researchbot specialist runs one research question with source evaluation and triangulation of major claims. See ResearchBot.

Voice

• Salesbot is a browser voice agent that runs a short sales discovery conversation; it is open to the public.
• VOX Presenter pairs a fast voice agent, which discusses a document aloud, with a second agent that writes a fuller analysis on screen. A session starts it with the voice sessions skill.

Around the work

• The goal map's sweep agent surveys running and recent agent sessions and records what each is working on and waiting for, which the Goal Map / Agent Orchestrator draws around the month's goal.
• A diagram agent takes a brief and returns a finished architecture diagram, using the C4 diagram skill, so the session that asked does not load the diagram rules itself.