Runtime
Hermes provides capability
Tools, memory, file access, scheduling, messaging, browser workflows, terminal workflows, and local automation.
Loop OS / AgentOps roster
Named capacity, not random bots.
A roster for agentic work: who coordinates, who specializes, what runs underneath, and where human judgment stays in the loop.
Operating model
Capability, coordination, labor, judgment.
The roster separates infrastructure from agent identity. Hermes is not a persona in the cast. It is the substrate. StraatBot coordinates. Specialists do domain work. Humans retain high-stakes judgment.
Four-part architecture
A coherent system needs more than prompts.
Orchestration
StraatBot coordinates
Capture, triage, routing, delegation, open-loop tracking, status updates, and escalation.
Specialization
Agents provide labor
Research, critique, prototypes, synthesis, planning, operations, and domain-specific drafting.
Judgment
Humans stay accountable
Approvals, commitments, publication, spending, permission changes, and final editorial standards.
The roster
Agents with scope, soul, stack, and boundaries.
Each role has a job, an operating temperament, expected outputs, and explicit approval gates. Infrastructure appears in the roster only to show what it enables — not as a persona.
Permission ladder
Autonomy gets safer when it is named.
Permission levels make the system inspectable: read-only, draft, internal write, reversible execution, external-with-approval, and manual-only work.
Usage and cost visibility
Consumption belongs near the roster — with room to become its own console.
The roster answers “who is this capacity?” The consumption layer answers “what is it costing, where is it happening, and when should the system slow down?” Put the summary here so every agent has an operating budget; graduate to a dedicated platform view when live billing, quotas, and alerts need daily management.
30-day tokens
231.3MAll Hermes sessions, including Telegram, CLI, cron, and TUI.
Primary model
gpt-5.5224.2M tokens across 230 sessions in the current 30-day window.
Highest channel
Telegram178.5M tokens; most visible place for approaching-limit warnings.
Tool calls
4,677Useful proxy for agent labor intensity and repeated workflow cost.
Operating recommendation
Keep roster and consumption connected, but do not collapse them.
Roster page: show each agent's default model, autonomy level, budget posture, and recent usage badge. This keeps agent capacity legible.
Consumption view: provide the finance/operator dashboard: spend by model, agent, surface, workflow, client/project, and alert band.
Rule of thumb: if the question is “who should do this work?” it belongs in the roster. If the question is “are we burning too much capacity?” it belongs in consumption.
Live instrumentation
Events to capture
- Agent
- Attribute every run to the named role that initiated or received the work: StraatBot, Scout, Sterling, design squad, or specialist.
- Model
- Break spend and quota pressure down by provider, model, input, output, cache/read/write tokens, and retry/fallback path.
- Surface
- Show where demand originates: Telegram, CLI, cron, dashboard, Kanban worker, or embedded client workspace.
- Outcome
- Pair consumption with business value: report drafted, issue closed, artifact created, message prepared, reminder completed.
Normal
Quiet baseline. Show trend, last 7 days, and top agents/models.
Watch
Badge the roster and summarize the drivers behind increased usage.
Review
Prompt for model-routing changes, batch limits, or approval before heavy runs.
Limit
Gate expensive agents, pause nonessential cron jobs, and require explicit approval.
Handoff flow
How work moves through the system.
Luis
Direction, taste, approval, and final commitments.
StraatBot
Orchestration, triage, routing, and escalation.
Specialists
Scout, Sterling, design squad, and domain roles.
Hermes
Runtime layer: tools, scripts, memory, messaging, files.
Truth
Human-readable notes, reports, queues, and operating views.
Ollin
The pulse that keeps loops moving.
Ollin is the motion layer of Loop OS: a cadence system for turning context into coordinated action.
Motion cycle
Capture → clarify → route → move → review → learn.
Where most tools store information, Ollin asks what should move next. It connects capture, context, cadence, review, and agent handoffs so important work does not remain trapped as notes, scattered chats, or forgotten intentions.
Bring the model to a team