Overview
One pane over the whole governed estate: agents, automations, guardrails, and policies — with live activity, cost, and posture. Everything here is created and run through the orchestrator or by hand, and every object links to its home.
Governance posture click any chip to jump to it
Recent activity unified trace log · click a row
| When | Kind | Name | Status | Cost |
|---|
Quick actions
Governed Chat
The assistant can really call every connector on the Integrations page, search the knowledge base, and assemble automations from conversation. Every input and output is guardrail-scanned.
Ask anything. Or assemble a workflow.
The assistant has live access to every connector, the knowledge base, and the automation builder. Everything is guardrail-scanned.
Automations
The North-style lifecycle: describe the goal in plain language, review and edit the approach in Plan mode, test and iterate, then publish to production with a schedule. Loops and branching keep the execution path auditable; every change is versioned.
1 Describe
Your automations
Agents
Agents provisioned by the orchestrator or built by hand: purpose, routing tier, tool access, and the guardrail profile each one runs under. Guardrail profiles are managed under Governance → Guardrails.
Registered agents
Deployments
One-click deploy for published agents. AMP Runtime serves the agent at a live, callable endpoint on this control plane. The other targets are deployment orchestration: AMP emits the real artifact (App Runner, container, edge, or a PR) for your runtime to run — the build-plane→runtime-plane handoff — while governance, traces, and FinOps keep flowing back here.
Deploy an agent published agents only (eval-gated)
Deployments
Knowledge Base
Real indexing: documents are chunked and served through BM25 retrieval. Chat and automation steps query this index through the knowledge connector. Upload .md, .txt, or .csv files.
Index a document
Try retrieval
Generate a starter document ✦ agent-assisted
Search results
Indexed documents
| Document | Chunks | Size | Indexed |
|---|
Integrations
Four integration surfaces — Sources, Destinations, Model providers, and MCP — plus the extras most platforms don't ship yet: Kafka streaming and Jira are working here today, and the whole connector registry is served as native MCP. Green badges are functional right now; test any of them live.
Invoke an action directly what the assistant does under the hood
Model Registry & Routing
Routing tiers map to real models from the configured provider. Steps declare a tier, never a hard-coded model, so the fleet upgrades itself. AMP is provider-agnostic: run agents on Anthropic's managed API, or on any OpenAI-compatible endpoint including fully local models.
Model provider
| Tier | Model | Est. $/MTok in / out | Agents on tier | Session spend | Used for |
|---|
Guardrails & Violation Evidence
A real detection engine runs on every chat turn and automation step: PII, secrets, and prompt-injection heuristics with matched spans and confidence. Violations below are from this session's actual traffic.
Test the engine
Active rules inspectable, not a black box
Guardrail profiles reusable bundles agents run under · manual or ✦ agent-assisted
Violations this session
| When | Source | Entities | Action | Redacted preview |
|---|
Monitoring Policies
Live policy evaluation against this session's actual run statistics. Create policies manually, or describe one in plain language and the orchestrator drafts it for you.
New policy manual or ✦ agent-assisted
Active & monitored policies
| Policy | Current | Threshold / rule | Status |
|---|
Trace Explorer
Every chat turn and automation run in this session, with step-level spans: connector calls, LLM steps, approvals — durations, tokens, and cost.
Traces click to open the waterfall
| When | Kind | Name | Status | Duration | Cost |
|---|
Smart Model Routing
The model is chosen per task, not per agent. Every chat turn, automation step, branch judge, agent ask and plan is classified (task class + complexity), scored against the fleet under a capability floor, and routed to the best fit — with the why recorded for every decision. Decisions join to runs and attributed outcomes, so value shows up next to the model that produced it.
Routing policy the tradeoff every decision is scored under
Model fleet capability profiles · live usage · attributed value
Task class → model distribution
Value by model from attributed outcomes, distributed by run-cost share
Try the router dry-run — classified and scored, nothing logged or billed
Routing decisions newest first · why this model
| When | Feature / agent | Task | Complexity | Model | Est cost | Why |
|---|
Cost & Token Metering
Real usage from this session, metered per trace: tokens in and out, estimated spend at configurable tier rates, and where the money went.
Platform intelligence vs workspace usage what gets rebilled vs passed through
Spend by agent from playground + runs
Spend by workflow
Spend by step type
Spend by kind
Spend by trace
| Trace | Kind | Tokens | Cost |
|---|
Solution Library
A curated, pre-populated catalog the orchestrator draws from: workflow blueprints, agent use cases, reusable guardrail profiles, governance policies, agent design patterns, and starter automations. Ask the orchestrator to "set me up" and it recommends and assembles the right ones for you.
Documentation
Living docs: guides on how the platform thinks, a complete API reference generated from the same manifest the server runs on (with a working try-it console and an OpenAPI export), and wiki pages you can write yourself — persisted with the rest of the tenant.
Self-Healing Remediation
AMP watches the running tenant for real problems — a connector that failed its health check, a stopped deployment, a workflow that errored, a spend or guardrail breach — and offers a playbook for each. Automatic steps take real action; anything consequential stops and waits for a named human.
Detected issues
Remediation runs auto steps + human gates
Tenants
Operator view. Each tenant is a fully isolated workspace: its own process, its own data, its own access code, spend ceiling and rate limit. Suspending a tenant stops its workspace immediately; deleting one archives its data. Access codes are shown once — copy them when they appear.
Create a tenant
Tenant estate isolated workspaces
Users & Organizations
Approve sign-up requests, invite people into an organization, and manage roles and seats. Sign-ups wait here until approved; invites activate the moment they are accepted. Each organization owns one isolated workspace and a fixed number of seats. Org admins manage their organization's people and seats; Members use the workspace; the Super admin runs the platform.
Invite someone
Organizations seats and plans
Users
Platform Analytics
The whole estate at a glance: who is on the platform, how much they use it, what each organization's workspace is spending, and everything that has happened recently.
Sign-ins per day last 14 days
API requests per day last 14 days
Organizations seats · activity · workspace spend
| Org | Plan | Seats | Members | Pending | Last activity | Spend | Platform AI billed | Status |
|---|
Recent activity platform event stream
Outcome Attribution
This is what makes value measured rather than estimated. Each workflow gets a rule that turns a business outcome into attributed dollars. Work-product outcomes attribute automatically when a production run succeeds; outcomes that depend on an external acceptance wait for a real signal, which any CRM, helpdesk or billing system can post to the inbound receiver.
Outcome rules
Inbound signal receiver wire a real system
Value by workflow
Value by model which models produced the attributed value
Recent signals
Attributed outcomes newest first
Workflow Dashboards
Every deployed blueprint gets a custom dashboard generated automatically — live agent status, run volume, operating cost, guardrail posture, and an Agentic Value Yield estimate (value delivered minus operating cost).