Selected Work
Agent Infrastructure 2026 Platform Suite

Locus + Atrium

Agents on the org chart.

Agentic AI works in demos. In production, it falls apart at the coordination layer — which agent did what, when, with whose data, on whose authority, with what cost.

Teams that try to scale past a single agent hit the same wall: operations. No visibility. No audit. No control. Impressive demos that can't survive contact with a real organization.

Rogue Agents is building the suite that solves it. Two products, one underlying model: agents as first-class operators, humans with real visibility and real control.

Atrium for individuals and small teams. Locus for enterprise. The same model underneath.

4 Layers Identity, audit, coordination, surface
Solo → Enterprise One growth path, no migration required
Plain text Fully auditable agent state, no black boxes

Make running a team of agents feel like running a team of people.

Most agentic AI was built to impress in a demo. You give it a goal, it executes, results appear. Clean. Impressive.

Production is different. In production, you need to know:

  • which agent made which decision, and when
  • what data it accessed, and on whose authority
  • where it escalated to a human, and what that human decided
  • what the full history of any agent looks like, reviewable in plain text

This is the operations problem — the gap between "agent that can do a task" and "agent infrastructure a real organization can actually run." Locus and Atrium exist to close that gap, from a solo operator's desk to an enterprise's org chart, with the same model of how agents work underneath.

Project Locus + Atrium
Category Agent Management Suite
Products Atrium · Locus
Built 2026
Status In Development
Humans in the loop.

Agents on the org chart.

Three commitments shaped the suite.

  1. 01 /
    Legibility over magic. Everything an agent does is auditable in plain text. No opaque state, no decisions you can't review later.
  2. 02 /
    Authority is explicit. Every agent has a defined scope. Widening it requires a human decision, recorded.
  3. 03 /
    The right tool for the size of the team. A solo operator needs a different surface than an enterprise — so we built two surfaces, sharing the same model underneath.
Primary / 02

Locus

Enterprise agent infrastructure. Multi-team, multi-tenant operations. Org-level authority and audit. Built for the organization where agents need to be managed the way employees are — with identity, accountability, and control at scale.

  • Multi-team, multi-tenant agent operations
  • Org-level authority, identity, and role-based access control
  • Integration with enterprise auth and existing infrastructure
  • Audit and compliance at organizational scale
  • Grows from Atrium without rebuilding agents, data, or workflows
The Suite

One Model, Two Surfaces

Atrium and Locus share the same underlying architecture. A team can start with Atrium and graduate to Locus without rebuilding their agents, their data, or their workflows. The suite grows with the team — solo to small to enterprise — without breaking the operator's mental model.

Available Now

Atrium ships to individual operators and small teams in summer 2026. Locus is being deployed in enterprise environments where agentic AI needs to scale past one team. For evaluation engagements, design partnerships, or enterprise rollouts, pitch us a project.

From a single operator's desk to an enterprise's org chart — one model of how agents work.

Two surfaces. Same substrate. Built to grow with the team. Four layers, from identity to surface.

01 / Identity & Authority

Agents are first-class operators, not session IDs.

Every agent in the suite has a persistent identity — a name, a history, a set of relationships, a defined scope of authority. Not a session that resets. Not a thread that disappears. An operator with a record.

Every agent's authority is explicitly bounded. Widening that scope requires a human decision, recorded in the audit trail. The trail spans the agent's full lifetime, not just the current session — so nothing disappears between conversations, and nothing is decided without attribution.

Persistent identity — name, history, relationships across every session
Bounded authority — every agent operates within an explicit, defined scope
Human-gated widening — scope changes require a recorded human decision
Lifetime audit — the trail begins at creation, not at the current session
AGENT // IDENTITY RECORD
AGENT NAMEResearch-01
CREATED2026-05-12
SESSIONS143
AUTHORITY SCOPEREAD + DRAFT
SCOPE WIDEN REQ.HUMAN DECISION
AUDIT TRAILLIFETIME
02 / Operations & Audit

Everything an agent does is reviewable in plain text.

Agent state lives in markdown — readable, diffable, version-controllable. No opaque internal state. Nothing the operator can't open and read. Every action is timestamped, attributed, and reviewable after the fact.

Replay and inspection are first-class operations, not afterthoughts. If an agent made a decision three weeks ago that now looks wrong, you can find it, read it, and understand exactly what happened. The audit isn't a compliance layer bolted on — it's structural.

Plain-text state — markdown files, readable and diffable by anyone on the team
Full attribution — every action timestamped and attributed to the agent or human that took it
Replay-ready — review any point in an agent's history, not just the current state
No black boxes — nothing lives in opaque internal state
AUDIT LOG // PLAIN TEXT // MARKDOWN
14:32  Research-01RETRIEVED DOC
14:33  Research-01DRAFTED BRIEF
14:35  Research-01ESCALATED →
14:40  NateAPPROVED
14:41  Research-01DELIVERED
03 / Coordination

Agents working with humans, and with each other, in real workflows.

Real agentic work isn't a single agent executing a single task. It's handoffs — agent to agent when the work changes character, agent to human when a decision requires authority or judgment beyond the agent's defined scope.

The coordination layer defines those handoff patterns, the escalation rules, and the shared context that flows across an agent team. Memory and state persist between agents. Nothing disappears at the boundary between one agent and the next.

Agent-to-agent handoffs — work passes cleanly between specialized agents
Escalation rules — explicit conditions that require human input before proceeding
Shared memory — context and state persist across the full agent team
Tool integration — coordinates with the systems the team already runs
COORDINATION // ACTIVE
Research-01
Writer-01
HANDOFF // AGENT-TO-AGENT
ESCALATION THRESHOLD REACHED
HUMAN // MUST APPROVE
04 / Surface

Atrium for individuals. Locus for enterprise. Same model underneath.

The two surfaces share the same underlying architecture. A team can start in Atrium — the desktop-based, cloud-backed management layer built for individual operators and small teams — and graduate to Locus without rebuilding their agents, their data, or their workflows.

The mental model is constant from solo to enterprise. Only the surface changes: from a single-operator desktop to a multi-team, multi-tenant organizational infrastructure with enterprise-grade identity, access, and audit baked in.

Atrium — desktop-based, cloud-backed, built for individuals and small teams
Locus — multi-team, multi-tenant, enterprise identity, RBAC, and audit at scale
No migration — start in Atrium, grow to Locus, keep the same agents
Constant mental model — solo to enterprise without breaking operator intuition
SURFACE // TWO DEPLOYMENTS
Atrium DESKTOP-BASED CLOUD-BACKED LOCAL-FIRST SPEED SOLO + SMALL TEAM
Locus MULTI-TEAM MULTI-TENANT ORG-LEVEL AUDIT ENTERPRISE RBAC
SAME MODEL // SAME AGENTS // NO MIGRATION

What makes the infrastructure hold up in production.

The differentiators here are architectural. Four commitments that shape every decision in the suite — and that explain why it's one product, not two.

  1. 01 /

    Plain-text state

    Anything an agent does is reviewable in a markdown file. No black boxes, no proprietary formats. State that any operator can open, read, and version-control.

  2. 02 /

    Bounded authority

    Agents act within explicit scopes. Authority widens only by human decision, and that decision is recorded. The system cannot grant itself more power.

  3. 03 /

    Persistent identity

    Agents have names, histories, relationships — not session IDs that reset. An agent that worked with your team last month still remembers last month.

  4. 04 /

    Two surfaces, one substrate

    Atrium and Locus run the same model of how agents work. Start small, grow large, don't migrate. The architecture doesn't change when the team does.

Built for production.
Deploying now.

The suite is in active development. Atrium ships to individual operators and small teams in summer 2026. Locus is being deployed in enterprise environments where agentic AI needs to scale past one team, with the operational visibility the enterprise requires.

The production gap in agentic AI is real. Most teams hit it around their second or third agent. The coordination problem isn't hard to name — it's just hard to build for, and almost no existing tooling was designed with it in mind from the start.

Available now under the Rogue Agents banner. For evaluation engagements, design partnerships, or enterprise rollouts: pitch us a project.

  • Atrium ships summer 2026 to individual operators and small teams
  • Locus in deployment for enterprise environments requiring org-scale agent operations
  • Shared architecture: start in Atrium, grow to Locus, rebuild nothing
  • Four architectural commitments: plain-text state, bounded authority, persistent identity, unified substrate
  • Available for evaluation engagements, design partnerships, and enterprise rollouts
What's next

Deploying agents at
scale?

Tell us where you are in the build. We'll tell you how the suite fits.

Pitch us a project
← Back to selected work