GhostFactory
AOS · Docs ← Back to the app

How the builder demo works

GhostFactory AOS Control Plane is where you would run a fleet of AI agents split into factories: who exists, what each run did and cost, which policy stopped it, and where a human has to sign. This is a clickable demo of it. The operations data is invented. The agent catalog is real: it comes from K-B, the GhostFactory knowledge base.

The catalog holds agents in teams, grouped into factories. of them have a running instance in the demo fleet.

What this is

The demo shows what operating an agent fleet could look like, from the catalog down to a single signed decision. It also holds itself to one rule: a number on a screen is derived from data underneath it, not typed into the page. Where two screens show the same fact, they read it from the same place.

How to read this page

Every screen section carries one or more tags. Notes call out controls and texts that promise more than the demo does.

live
Computed or written in your browser, and read by other screens: simulations, the ledger, Shield, HITL Queue, Lab.
K-B
Real catalog data from K-B: agents, teams, factories, contracts, model recommendations.
design
A screen or panel with invented data. You can browse, filter and open details; nothing changes underneath.
plan
A control or a claim the demo does not implement. Clicking it does nothing, or only shows a toast.

The demo keeps its state in your browser, and there is no reset button. See What your browser keeps.

Catalog and fleet

Two different things, and the screens label which one they show.

Catalog
Who exists. orchestrators and members from K-B, each with a role, a team, a factory, contracts and a recommended model.
Fleet
What runs. Instances with status, load, uptime and task counts. The fleet is invented; sealing an agent in the Agent Builder adds an instance until the next reload.

An agent in the catalog with no instance is a state, not a gap. The agent card says so: "The agent is in the catalog, but the fleet has not deployed it." Agents lists it under Catalog, Factories counts catalog and deployed separately, and Topology marks a team with nothing deployed.

Factories, teams and agents

A team is an orchestrator and its members, as K-B defines them. A factory pairs K-B teams into one unit with its own graph in the Editor. Every agent carries its K-B contracts: the skill it performs, what it requires, what is optional, what it returns, and its token budget. K-B also recommends a model for each agent and lists models to avoid; the agent card compares that with the model the instance runs.

The Topology map draws teams as boxes. Not every team has a box on the map yet; the full catalog is on Agents.

Runs, spans and the ledger

You start a run in the Editor under Simulate. Each step of the run is a span, and each span is appended as an event to the ledger. When the run ends, its events are sealed into a block. A block stores the SHA-256 hash of its contents, the hash of the previous block and a merkle root, so the chain is append-only: changing an old entry breaks every hash after it.

Overview, Runs & Traces, Cost & Tokens, HITL Queue, Ledger / Audit and Lab read the ledger. A run you simulate shows up in all of them, next to the invented history they start with.

Verify Integrity on Ledger / Audit recomputes every block hash. Tamper edits a sealed entry so that Verify has something to catch.

Warning

Tamper is saved with the ledger. The chain stays broken in this browser, across reloads, until you clear the site data.

Shield and human sign-off

Shield rules are predicates the simulation checks. Task rules that refuse outright are checked before a run starts; the rest at the policy checkpoint, once the run's real cost is known. Step rules are checked before each step that reads or writes something. A rule does one of two things.

escalate
The run parks and waits for a signature. The request appears in HITL Queue and the sidebar shows a count.
block
The step is refused. Some rules re-route it (a write to a production path goes to staging instead); otherwise the run ends.

Examples you can try: the preset "Write prod migration" parks on R.01, a destructive operation. Switch R.01 off on Shield Rules and the same run parks on R.03 instead, because it costs more than its factory's per-task ceiling. A task that contains an email address is refused by R.08 before anything is dispatched.

Approve resumes the run, which then finishes and seals its block. Deny ends the run; the remaining steps are skipped and the block is sealed with the denial. Both decisions are ledger events signed by you.

What a run costs

An agent loop re-sends the whole conversation on every model call. A run therefore pays mostly for input, not for the text the model generates. A tool result, such as a file the agent read or a test transcript, joins the conversation and is paid for again on every later call.

The prompt cache softens this: the part of the conversation an earlier call already sent is read from cache at a fraction of the input rate, and writing to the cache costs a little more than plain input. The simulation prices planning and generation on the orchestrator model and research and review on the cheap tier. Rates come only from the model registry. The run drawer on Cost & Tokens splits a run into fresh input, cache read and output.

Models and routing

The model registry lists models. It is built from the public catalog at models.dev during maintenance, not fetched while you click, so prices are as of the last update. Models are grouped by their role in an agent system: Orchestrator, Worker, Cheap, Specialist, Local, Embedding and Deterministic.

Each routing role has a routing chain: a primary model, a failover at a different vendor, and an escalation. You see it in three places:

What your browser keeps

The demo stores its state in this browser's local storage for this site:

gf_ledger
the ledger: events and sealed blocks
gf_shield_rules
which Shield rules are switched on
gf_lab
Lab jobs, configs and the open tab
gf_data_mode
Seeded or First-run
gf_theme
Light or Dark; this page follows it
gf_screen
the last screen, reopened when you come back
gf_ov_range
the range picked on Overview
gf:cmdk:recent:v2
recent command palette entries

Agents sealed in the Agent Builder and factories sealed in the Factory Builder live only in memory and are gone after a reload.

planNo reset

There is no reset button. To start over, clear this site's data in your browser settings. That also repairs a chain broken by Tamper.

01 Control Plane

Overview livedesign

What did the fleet do since you left? Runs completed and failed, money spent, sign-offs waiting, trends and incidents. The figures are computed from the ledger and from an invented run history; the header counts agents and factories from the fleet.

planDoes nothing

View buttons and the incident rows.

Factories K-Bdesign

Which factories exist and what is in them? Factories and their teams come from the catalog, with catalog and deployed counts per factory. Throughput figures are invented.

planDoes nothing

Scale and Quiesce.

Topology K-Bdesign

How do teams connect? A company map where each box is a K-B team, colored by what the fleet has deployed. Channels, call counts and SLAs are invented. Not every team has a box on the map yet.

Agents K-Bdesign

Who exists and who runs? Two lists: Deployed (fleet instances, with invented load and uptime) and Catalog (every agent from K-B). See Catalog and fleet.

Editor livedesign

What does a factory look like inside, and what happens when it runs? A node graph of the factory with a library of Agents, Skills, Models, RAG, Rules, Packages and Flow. The graphs are designed examples; the Models tab reads the model registry.

Build, Route, Validate
Build is the plain editing view; Route animates the routing paths; Validate marks findings on the graph.
Deploy
Opens a dry-run diff of the graph.
Simulate
Opens the run drawer. A preset only fills in the task; Run starts it. Steps stream into the trace, every step lands in the ledger, Shield checks run as described in Shield and human sign-off, and the finished run is sealed into a block.

Library cards, the + on a card and the Graph, Flow and Schema views change the canvas.

planToast only

Deploy to Staging in the dry-run dialog only shows a toast saying that nothing was deployed. The options in the dialog change nothing.

Agent Builder K-Bdesign

How would you define a new agent? Pick a template, then walk the steps from identity to deploy. The templates are a few generic workers plus every agent in the catalog; a catalog template brings its K-B role, contracts and a system prompt composed from them. The model picker follows the routing chain of the chosen role.

Seal & Deploy adds the agent to the fleet and opens the Editor with its node. It writes nothing to the ledger or the ADR Vault. The instance lives in memory and is gone after a reload.

Factory Builder design

How would you design a new factory? Pick a topology (ROMA Planner, Pipeline, Map-Reduce or Review loop), then go through Identity, Topology, Squad, Guards and Review.

Runs & Traces livedesign

What did a run do, step by step? A list of runs (invented history, with your simulated runs on top) and a Trace Inspector that shows the spans of the selected run as a waterfall.

planDoes nothing

The OK, Warn, Err and HITL pills only change their look; the list is not filtered. Also All, the factory and time pickers, Export CSV, Replay, View in Editor, Open ledger block, Flag and As eval set.

Cost & Tokens livedesign

Where does the money go? Spend per agent, grouped into factories by the catalog, priced with the rates of the model registry. Simulated runs join the list of the most expensive runs. Budgets and month-to-date billing are invented.

planDoes nothing

Resolve and Edit.

Inbox design

What needs attention right now? One feed of HITL requests, Shield blocks, quarantine flags, cost alerts and connector errors. The feed is invented.

planDoes nothing

Severity, source, team and time filters; Unread, All, Mentions and Subscribed; Mark all read; Filters; Approve refund, Edit amount & approve, Reject & send to support; Edit.

Schedules design

What starts runs without a human? Cron, webhook and event triggers. All invented.

planDoes nothing

The on/off switches change their look only. Nothing is saved or started.

Deployments design

Which version runs where, and is the canary safe? Versions, traffic split, eval gates, a deploy log and rollbacks. All invented.

planDoes nothing or toast only

Hold canary, Rollback and Promote → 50% do nothing. Promote when soak passes, Open incident, View diff, Forward to ADR Vault and Acknowledge only show a toast. An adapter promoted from the Lab does not appear here.

Evals & Replay design

Does a new version break anything? Golden sets, adversarial cases and red-team suites, a pass/fail matrix, a regression diff and production replay. All invented.

planDoes nothing

Open ADR, Add to red-team suite, Re-prompt, Skip case, the replay pickers, Run replay, Save as suite and Cancel.

02 Data & Knowledge

RAG Stores design

What can agents look things up in? Five stores: a golden-set YAML file used as a fast path, and vector stores. All invented.

planDoes nothing

Reindex, Export · JSONL and Pin to Retrieval.

Skill Libraries K-Bdesign

What can agents do, and with which tools? Two layers, and they are different things:

Skill Libraries
The skill contracts from K-B, per team: triggers, input and output shape, constraints.
Tools / MCP
The tools of this demo that the Editor places on the canvas, such as reading a file or querying a store. They are not K-B skills.

Switch layers at the top and filter by team.

Knowledge Graph K-Bdesign

What depends on what? Entities and their relations: factories, agents, skill libraries, tools, stores, ADRs, eval suites and incidents. The graph shows a hand-picked selection of factories and agents, not the whole catalog. Their nodes read role, team, contracts and model from K-B; runtime figures on them are invented and labeled as such.

03 Decisions & Policy

ADR Vault design

Why is the system built the way it is? Architecture decision records with their status. The records are invented.

planDoes nothing

Seal & Accept, Request changes and Add reviewer.

Shield Rules live

What is an agent not allowed to do? Eight rules, R.01 to R.08, each with its scope (task or step) and action (escalate or block), and how often it fired this session. The simulation checks exactly these rules; see Shield and human sign-off.

HITL Queue live

What is waiting for a human? Pending sign-offs from the ledger, with Approve and Deny. When the queue is empty, a link takes you to the Editor to create one. What Approve and Deny do to the run is described in Shield and human sign-off.

Recent Decisions lists your decisions, marked LIVE, above an invented history. The totals in the header add your decisions to fixed numbers.

04 Integration

Connectors design

Which outside services can agents reach? Eight integrations with their status. All invented.

planDoes nothing

Configure, Rotate Key and Disable.

05 Safety

Quarantine design

Which agents are isolated, and why? Tripped circuit breakers, isolated agents and resolved cases. All invented.

planDoes nothing

Inspect Stack, Release and Terminate.

Ledger / Audit livedesign

Can the record be trusted? The chain of sealed blocks: an invented seed history plus every block sealed in this browser. See Runs, spans and the ledger.

planDoes nothing

Export Chain · CSV, Download Compliance Bundle and Export Block · JSON in the block drawer.

Reliability design

Is the platform meeting its targets? SLOs, error budgets, active incidents and post-mortems. All invented.

planDoes nothing

Acknowledge, Page secondary, Escalate, Service health → and Edit post-mortem.

06 Lab

Lab livedesign

Can the fleet learn from its own record? Datasets are definitions over the ledger: sealed blocks, run spans and HITL decisions as preferences. Their counts grow as you simulate and sign.

Datasets
What each dataset selects; Tune → opens Tuning with that dataset picked.
Tuning
Create job and Start training. Training paces through its epochs and draws a loss curve.
Configs
Edit a queued job's settings; Save config records a new revision.
Eval Gate
Compares the result against the eval suites; Promote to Deployments is enabled when the gate passes.

Creating a job, training, saving a config and promoting are all ledger events; they are not sealed into blocks. The open tab and your jobs are remembered.

planSimulated
  • No model is trained and no adapter is produced. The loss curve and the eval pass rates are simulated; they are deterministic, so the same job gives the same numbers.
  • Promote to Deployments opens Deployments, which does not show the promoted adapter.
Around the screens

Command palette design

⌘K or Ctrl+K opens a searchable list: Recent, Navigate, Actions, Simulate, Factories, Agents, Runs and ADRs. The list is fixed; it is not built from the sidebar or the catalog.

planNot wired to the demo
  • Export compliance bundle and Pause all factories only show a toast. Toggle theme reports that it is not available; use Light and Dark in the top bar.
  • Simulate entries raise a scripted incident in a strip across the top of the screen. It does not reach HITL Queue or the ledger. Approve all and Deny all pending HITL act on those scripted incidents only, not on sign-offs in HITL Queue.

Atelier guide design

The button marked A at the bottom right opens a chat panel. It answers from a script: your question is matched against keywords and the guide replies with prepared text and figures, plus a link to the screen that shows the topic. A question it does not recognize gets "I don't have a tool wired for that yet." The header and the welcome text say that the answers are scripted.

planScripted

No model answers and nothing you type leaves your browser. The tool calls and the model name the panel shows are part of the script, and the figures in the answers do not follow the screens.

Top bar and rail

Seeded / First-run
First-run shows the empty state of seven screens: HITL Queue, Quarantine, Ledger / Audit, ADR Vault, RAG Stores, Skill Libraries and Connectors. The others stay as they are. Remembered.
Light / Dark
The theme. Remembered, and this page follows it.
Emergency Stop plan
Does nothing.
+ Spawn Factory
Opens Factories.
Play Incident
A scripted walkthrough of about 18 seconds: a destructive command is caught by Shield, escalated, denied in HITL Queue, isolated in Quarantine and sealed in the Ledger. It moves between those screens with captions; arrow keys step, Space pauses, Esc quits. It writes nothing to the ledger.
How this demo works
The link under the logo opens this page.

The footer of the sidebar is static text.

Reference

Glossary

Factory
A unit that pairs K-B teams and has its own graph. More
Team
An orchestrator and its members, as K-B defines them.
Orchestrator
The agent that leads a team and hands work to members.
Member
An agent in a team that does one kind of work and reports back to the orchestrator.
Catalog
Every agent K-B defines. More
Deployed / on the shelf
An agent with a running instance in the fleet, or a catalog agent without one.
Contract
A K-B skill an agent performs: what it requires, what is optional, what it returns, its budget.
Affinity
The model K-B recommends for an agent, and the models it advises against.
Primary / failover / escalation
The links of a routing chain: the default model, a fallback at another vendor, and a stronger model to escalate to. More
Model registry
The list of models with their roles and rates, updated from models.dev.
Run
One execution of a task by a factory; in the demo, a simulation.
Span
One step of a run, with its duration, tokens and cost.
Trace
All spans of a run, shown as a waterfall.
Ledger
The append-only record of everything that happened. More
Block / sealed
A group of ledger events closed with a hash that chains it to the previous block.
Genesis
The start of the chain, shown on Ledger / Audit.
Verify Integrity
Recomputes every block hash and finds the first broken block.
Tamper
Deliberately edits a sealed entry so Verify can catch it.
HITL
Human in the loop: a run waits for a person to approve or deny.
Shield
The policy rules every run is checked against. A rule escalates or blocks. More
Ceiling
The per-task spend limit of a factory; rule R.03 escalates a run past it.
ROMA Planner
A factory topology where a planner fans tasks out to a squad of workers.
Arbiter / Worker
A deterministic router without a model, and an agent that does the work.
ADR
Architecture decision record.
RAG store
A store agents retrieve context from.
Golden set
Known questions with known answers, used as a fast path and as an eval suite.
Skill Libraries vs Tools / MCP
K-B skill contracts versus the demo's own tools. More
Canary / Soak
A new version on a small share of traffic, and the waiting period before it is promoted.
Eval gate
A check that must pass before a version or adapter is promoted.
Circuit breaker
Stops an agent that misbehaves, such as one stuck in a loop or over its token limit.
Quarantine
Where an agent goes when its breaker trips: no traffic in, no traffic out.
SLO / Error budget
A reliability target, and how much failure is left before it is missed.
Connector
An integration with an outside service.
Seeded / First-run
Sample data or the empty states a new installation would show.
Cache read
Input tokens served from the prompt cache at a fraction of the input rate. More

Limits of the demo