⌂ Console
Organizations
This server runs one organization: its flows, agents, decisions, playbook and history. Set it up with the wizard, then open it to manage it.
Loading organizations…
Step 1 of 8
Choose a starting point
Pick a pre-built domain template or start from scratch. Templates pre-fill decision types, actions, governance rules, and ROI defaults.
Step 2 of 8
Organization identity
The organization ID keys everything this server stores. It is set where the server runs (ACTIONGRAPH_ORG).
Set by ACTIONGRAPH_ORG. One organization per server.
🎨
AppearanceStep 3 of 8
Decision types
Define what kinds of decisions this system will make. Each type has its own actions, context schema, and reward criteria — the RL engine learns independently per type.
Step 4 of 8
Intelligence engine
How each decision works, start to finish. Set which past cases count as similar and how far agents may go alone — everything else has a sensible default.
🧭 How each decision works
high riskmediumlow
⚑ has a rule✎ you can change it here
Decision playbook
The document ActionGraph stores and your agents read. Edit it in the last step.
▸ Advanced — legacy engine: recommendation tiers and scoring For systems created before the playbook. Agents using the playbook are not affected.
🎚
Recommendation Tier
Every recommendation runs one scorer and one learner — the decision kernel. This is where you choose
how much of it a decision type is allowed to use. Each tier adds one evidence source to the one
before it. Lower tiers are not degraded; they are what a decision type should use when the evidence above
it does not exist or must not be trusted.
On Precedents only,
policy_engine and
RLPolicyState are never queried. On + Policy rules,
RLPolicyState is never queried — so a recommendation there traces entirely to rules a human wrote.
System default
Applies to any decision type without an explicit tier below.
Per decision type
🔒 Cap is a ceiling, not a lock: an agent may request a
lower tier but never a higher one, and the refusal is written to the audit trail. On a
Precedents cap, any recommendation an agent claims at
decide-time is stripped as well, since that tier produces none.
⚡
Scoring Algorithm
Choose how actions are ranked. Both modes read the same decision ledger — switching
does not require migration or retraining.
System default
🧠
LLM Augmentation
After the kernel scores and ranks actions (pure math), an LLM can generate human-readable
rationales explaining why each recommendation ranks where it does.
The LLM never influences scoring — it only explains the result.
▸ Advanced — legacy engine: object model and scope keys Superseded by the playbook (situation fields replace scope keys). Kept for older systems.
🌐
Legacy object model (not enforced)This older model is not the playbook ActionGraph enforces. Decisions, rules and outcomes are set in the playbook (see the last step).
Define object types (entities like PullRequest, Service),
link types (relationships like authored_by, depends_on), and
action types (with preconditions and effects).
Agents register entities via
actiongraph_register_entity — the resolver
walks the graph to build rich context for scoring.
🔑
Context Scope Keys
Scope keys let the legacy engine learn separate Q-values per context value — e.g. scoping
code_review by repo learns a distinct policy per repository.
For the playbook, list the field under Similar cases are matched on above instead.
▸ Advanced — legacy engine: retrieval and RL tuning Used by the /rl endpoints and older systems, not by agents on the playbook. Defaults are sensible.
🤖
LLM Assist
Optional. Used for knowledge-graph extraction and chat. It never scores or ranks actions —
the kernel does that. Leaving this unset costs you nothing on the recommend path.
🔍
RAG Semantic Search
TF-IDF embedding similarity (128 dimensions, cosine, zero dependencies) finds the most contextually similar past decisions and returns raw precedents so the agent can judge whether a prior decision applies.
Engine specs
Embedding: TF-IDF Hashing Trick
Dimensions: 128
Similarity: Cosine (L2 normalised)
Dependencies: Zero — pure Python
Configuration
Minimum cosine similarity
📈
RL Strategy
This selects the scorer used by the advanced RL endpoints (
/rl/*, A/B experiments,
Tier 3 analysis). It does not change /agent/recommend, which uses the
kernel’s blended scorer. Four strategies are implemented — anything else falls back to UCB1.
Bandit Algorithms — Fast Online Learning
✓
UCB1 + EMA
Upper Confidence Bound with Exponential Moving Average. Best general-purpose default. Fast convergence, optimistic exploration.
✓
Epsilon-Greedy
Exploits best action (1-ε) of the time, explores randomly ε of the time. Simple, predictable, high-volume decisions.
✓
CQL (Conservative Q-Learning)
Penalises out-of-distribution actions. Best for mission-critical & compliance-heavy domains where safety > exploration.
Contextual — Feature-Aware Learning
✓
LinUCB (Contextual Bandits)
Ridge regression with exploration bonus. Learns feature weights from context dimensions. Best when context fields drive decisions.
⚙️
Strategy ParametersFor epsilon-greedy. 0 = pure exploit, 1 = pure explore.
Half-life for reward decay. Recent decisions weighted higher.
Exploration bonus for LinUCB. Higher = more exploration.
Minimum data before scoped Q-values are trusted over global.
Step 5 of 8
Action Rank
How each decision type ranks its options, from the evidence of similar past cases, and the track record an agent needs before it may ask for approval or act alone.
Step 6 of 8
Ontology
What your decisions are about. Each decision type acts on a subject — a Claim, a Shipment — and each option is an action on it, with a risk and the system that carries it out. The playbook is built from this.
Step 7 of 8
Governance & ROI
Define safety guardrails and map decision outcomes to business value. Governance policies control when human approval is required.
🛡️
Governance Policies💰
ROI Configuration
Map each decision type to monetary value. Powers the ROI dashboard and quantifies RL improvement vs. baseline.
🔗
IntegrationsWebhooks
External event receiver
GitHub
PR events, deployments
Slack
Notifications & approvals
Jira
Issue & sprint tracking
PagerDuty
Incident escalation
REST API
Custom polling connector
Step 8 of 8 — Final
Review & deploy
Everything correct? Set up your ActionGraph organization. All settings are editable from the dashboard after deployment.
📦
SDK Quick Start🐳
Environment VariablesOrganization set up
Your ActionGraph organization is ready to record decisions.
🚀
Next steps1. Give your agents the agent key and connect them over MCP:
POST /mcp2. Agents ask before they act:
POST /v1/recommend, then POST /v1/record3. Report outcomes so the track record builds
4. Approve what waits on a person on the Approvals tab
5. Watch outcomes and autonomy on the organization's dashboard