---
name: "core-fleet-roster"
type: core
license: CC BY-NC-SA 4.0
description: Index of the os.Claude Layer 3 fleet — the functional agents and the vertical operators, their labels, roles, and profile files. Read to see which agent or operator owns what, then open the matching profile (agent-relay, agent-forge, operator-bara, operator-bottrader, and so on) for the full mission and remit. Also holds the fleet's open naming questions, including the bot.Trader layer names that collide with names already in service here.
---

# Agent roster

The Layer 3 fleet that runs os.Claude. Two classes live here (see `core-operating-model`, Layer 3): **functional agents**, each owning one capability across every vertical, and **operators**, each owning one venture end-to-end and commanding the functional fleet. Each is a named, accountable role; open its profile for the full mission, remit, and composition. Functional agents carry two name forms — a **written** form for prose and an all-caps **branded** form that is also the Linear `agent` label value. Operators carry a lowercase `operator-<venture>` value in the separate Linear `operator` label group — ownership of the outcome, set independently of the `agent` executor group since 2026-09-03 (APP-1103) — not a branded one. Only the named functional fleet is branded; `human`, the `operator-*` values, and other-model values stay lowercase.

## Functional agents

| Agent | Label | Role | Profile |
| -- | -- | -- | -- |
| **Relay** | `R3-LAY` | Delivery management — routes, unblocks, holds the merge gate | `agent-relay` |
| **Forge** | `F0-RGE` | Implementation — builds and opens a PR, never merges | `agent-forge` |
| **Prism** | `PR-1SM` | Review — independent quality verdict; clears the merge | `agent-prism` |
| **Scout** | `SC-0UT` | Pipeline — captures people and opportunities, kept clean | `agent-scout` |
| **Pulse** | `PUL-5E` | Ops — system health and self-improvement; propose-only | `agent-pulse` |
| **Atlas** *(one slice live)* | `ATL-4S` | Planning — goals, prioritisation, cycle commitment; one slice live — cycle review & commit proposal (propose-only on cycle membership), the rest future | `agent-atlas` |
| **Sonar** *(commercial)* | `SO-N4R` | Design strategy, discovery & service design — upstream of Pixel and Reach | `agent-sonar` |
| **Pixel** *(commercial)* | `P1-XEL` | Brand and design system | `agent-pixel` |
| **Reach** *(commercial)* | `R3-ACH` | Marketing and go-to-market | `agent-reach` |
| **Draft** *(commercial)* | `DR-4FT` | Writing: copy, UX writing, headlines and long-form drafting; holds the writing standard | `agent-draft` |

The Linear `agent` label group enumerates these same values; `convention-linear` (*Labels*) carries that enumeration and must agree with this table — reconcile both when the fleet changes.

## Operators

Vertical roles, one per venture, that own an outcome and command the functional fleet within a granted mandate. The slot can be filled by the Owner, another human, or an AI — the same contract. The tier above is the **Owner** (Aled), who sets each mandate and gates the hard anchors (a strategic pivot, budget allocation).

| Operator | Label | Vertical | Profile |
| -- | -- | -- | -- |
| **Operator, BARA** | `operator-bara` | BARA — relaunch, weekly loop | `operator-bara` |
| **Operator, bot.Trader** | `operator-bottrader` | bot.Trader — **Orchestrator** of the vertical's four-layer model; two mandates, weekly loop: **A** the platform and its validation gauntlet · **B** the bot.Trader-B goal-and-guardrails agent | `operator-bottrader` |
| **Operator, Apps** | `operator-apps` | Apps accelerator — a **portfolio** (careerOS, app.fitness, Luna, os.Claude, GTD); allocates finite fleet capacity across bets, weekly loop; composes `skill-portfolio-allocation` | `operator-apps` |
| **Operator, A1** | `operator-a1` | A1 — the Association of One practice plus its live client engagements as sub-mandates, weekly loop | `operator-a1` |
| **Operator, Mr Pritchard** | `operator-mrp` | Mr Pritchard editorial — build one connected design-leadership position in public; Reach owns the pipeline, Pixel the visuals; weekly loop; AI-filled | `operator-mrp` |

**One row is one operator, not one mandate.** bot.Trader is the first venture where a single operator holds **two** mandates — each with its own goal, envelope, and reporting initiative, and neither a sub-clause of the other (APP-541, 2026-07-15). So read the Vertical column as what an operator owns, which may be more than one thing: where it names two mandates, they are held, assessed, and reported **separately and never netted** into a single "the venture is on track". A second operator for Path B is the named remedy if and when B matures; until then the separation is the operator's to hold, and `operator-bottrader` carries the separation rule in full.

**A row may also name a layer the operator occupies inside its own vertical.** bot.Trader settled into a **four-layer model** on 2026-08-09 — Scout (the bot.Trader-B agent, which explores and generates hypotheses), Researcher (the desk gate and the empirical gauntlet), **Orchestrator** (this operator), and Trader (deterministic execution of a promoted strategy, one instance per strategy). Flow is one-way: Scout → Orchestrator → Researcher → Owner gate → Trader. As Orchestrator the operator is the **joining function** — it routes harvested hypotheses into Path A's candidate queue, holds **aggregate exposure and portfolio drawdown** across per-strategy accounts that no per-account kill switch can see, and sets emphasis across promoted strategies. That work belonged to nobody until it was named (APP-749; architecture in APP-750); `operator-bottrader` carries it in full.

Two things this does **not** change. The four layers are a description of one vertical's internals, **not a new class of Layer 3 role** — only Orchestrator is a role at all, and it is this operator wearing its own hat. And the pipeline framing describes how work moves, not how it is scored: **the two mandates stand exactly as APP-541 granted them** until the Owner rules otherwise, held and reported separately. Whether the granted structure should follow the flow is an open Owner decision, named in `operator-bottrader`.

**The role is past its single-instance trial.** bot.Trader is the second live operator, so the shape is no longer under test on BARA alone. **Apps, A1 and Mr Pritchard now have operators** (`operator-apps`, `operator-a1`, `operator-mrp`; APP-578). `operator-mrp` is AI-filled with Aled the Owner (APP-578 decision 8), its mandate read from the Mr P charter set 2026-08-11 (MRP-18). The Apps *bets* — careerOS, app.fitness, Luna, os.Claude, GTD — do **not** each get an operator: a bet earns one by **graduating**, which is earning an operator (not a team — the plan's team limit is full), the Owner's call case by case (`core-operating-model`, *Graduation = earning an operator*).

Each operator's label value lives in the Linear **`operator` group** — its own group, not the `agent` executor group, since the Owner re-parented the five values on 2026-09-03 so that who executes and which vertical owns the outcome are two independent facts on a ticket (`convention-linear`, *Labels* and *Operator attribution*; APP-1103). That enumeration must agree with this table — reconcile both when a venture gains an operator. All five values in the table are live in the group. The Orchestrator framing changes no label and no enumeration: it describes what the existing `operator-bottrader` value covers.

## Open — a naming collision in the fleet vocabulary

Recorded here because the roster is where the fleet's names live, and a collision tolerated silently becomes two sessions meaning different things by the same word. **This is a flag, not a resolution — nothing is renamed, and no rename is proposed as settled. The Owner decides.**

The bot.Trader layer names introduced on 2026-08-09 collide with names already in service:

* **Scout** is the functional agent `SC-0UT` in the table above — pipeline, people and opportunities — and is also bot.Trader's layer 1, the caged bot.Trader-B agent. The two have nothing to do with each other.
* **Trader** is bot.Trader's layer 4, deterministic execution code. In earlier bot.Trader material it named an execution-only *agent* role paired with an **Optimiser**, and informally "bot A". Neither of those roles exists as a Layer 3 agent — the optimiser's nature made it a skill, and execution is platform code (APP-541, classify-by-nature) — so the words are loose in the estate rather than owned.

The question for the Owner: whether the roster adopts the four-layer vocabulary as fleet vocabulary at all, and if so how the two senses of Scout, and the residue of Trader and Optimiser, are separated. Until that is ruled, the two vocabularies are **disjoint in scope** — a layer name means something only inside bot.Trader, and the fleet names in the tables above are unchanged — so anything written about bot.Trader should make clear which sense it means rather than relying on the reader to infer it.

**The collision has now produced a wrong call during an incident, and it did so through a deployed artefact name rather than a conceptual one.** On 2026-09-14, mid-outage, a session reading `assoc-one/bot-trader`'s scheduled workflows — *Funding capture*, *CI*, *Scout cycle* — concluded that Mandate B's daily decision cycle was not deployed at all, and told the Owner that `operator-bottrader` might be reporting against a cycle that did not exist. `scout-cycle.yml` **is** the decision cycle: its own cron comments name OANDA's 17:00 New York close, `MARKET_HALTED` refusals, positions carried over a weekend, and `runCycle`. A reader who knows the fleet reads *Scout* and sees the pipeline agent or the hypothesis layer, and so does not see a decision cycle. While the vocabulary question stays open, **the mapping is stated wherever a deployed artefact carries a layer name**: in bot.Trader, `scout-cycle.yml` / *Scout cycle* is the Trader-layer daily decision cycle — not Scout the fleet agent (`SC-0UT`), and not an exploration pass. A deployed name is what a human reads at 11am during an incident, so it weighs more in the Owner's decision than the conceptual names; and renaming the workflow changes the check name, which sequences behind any required-status-check configuration (APP-1451). (APP-1453.)

## The assistant tier

A third class, and a **tier** rather than a peer: it serves the **Owner** (Aled) directly rather than a capability across every vertical (functional) or a venture end-to-end (operator), and it sits **above** both — it may direct an agent or an operator, Atlas and Relay included, where the direction serves a stated goal, secures clarity, or was asked for directly. The tier is defined once in `core-operating-model` (*Layer 3 — Agents and operators*, *The assistant*) and is not restated here; where this section and that one disagree, that one wins. On time it still sits **downstream of Atlas**: Atlas commits the cycle, the assistant turns the Owner's slice of it into booked time — direction authority and that sequence are different things and neither reads on the other. (Owner ruling, Aled 2026-09-29 — APP-1848.)

| Agent | Label | Role | Profile |
| -- | -- | -- | -- |
| **Aide** *(MVP)* | `A1-DE` | Personal chief-of-staff — sweeps the committed cycle for items needing Aled, sizes them (15m/30m/1h/2h; >2h → Relay to split), and books calendar holds in his hold windows (propose-and-confirm) | `assistant-aide` |

Aide reads Linear (the upcoming cycle only, read-only) and writes only Aled's calendar; her capacity signal goes to Atlas and oversized items route to Relay. She is a personal / Owner-serving role, distinct from both the delivery fleet and the vertical operators — position her label in the `agent` group accordingly. The MVP composes `skill-attention-sweep`, `skill-time-plan` (the `skill-operator-time` rename, APP-431, renamed again by `APP-2652`) and `skill-time-block`; the `assistant-aide` profile is authored. Prerequisites from the planning-and-time layer (epic APP-428): the `work.Focus` recurring event (APP-426, declared as `work.Hold`) and cycles enabled (APP-432) are **done**, so the MVP is buildable and testable against a hand-committed cycle now; the **automated** Atlas→Aide handoff additionally wants the initiative KPI baseline (APP-433) so Atlas's cycle-planning (APP-429) can run. The Linear `agent` group already carries `A1-DE` as a live value — this entry documents what routes to it, and `convention-linear` (*Labels*) must agree.

**System rule:** anything blocked is assigned to the operator and marked Urgent — whichever agent hits the blocker moves it to Blocked, assigns the operator, states the cause, and never jams its own leg. Anything beyond an operator's mandate — a strategic pivot, a budget-envelope breach, a goal-set change — is escalated to the Owner with a recommendation, never actioned unilaterally.
