Crescendo Lab
繁體中文Book a demo

Agent Lab Platform · Security & Governance

Give every agent a clear identity, permission boundary, and line of accountability.

Agent Lab Platform brings identity, policy, approval, usage, budget, audit, and runtime environments into one control plane. Governance is present before an agent goes live, not added after.

Explore the governance modelDiscuss enterprise rollout

Want to explore first? Sign up for Agent Lab and open a workspace.

Agent Lab Platform

A secure, governed, and auditable enterprise agent platform

The four products share workspace, identity, policy, approval, usage, and audit. Every agent follows the same enterprise rules from entry and decision to connection and execution.

effective = user ∩ assistant ∩ toolUser's own permissionwhat they already see in ERPAssistant's capabilitywhich tools it was grantedTool's own permissionwhat the credential allowsEffective access

Delegated identity:
assistants act as the person asking

For every tool call, the platform computes an intersection: what this person can already do in the downstream system, what this assistant was granted, and what the tool itself allows. It runs only if all three agree. Your existing permissions are the rules.

  • Knowledge retrieval also runs as the asker, so documents they cannot see are never cited
  • A connector can be a shared organization connection or the member's own authorization
  • Permission is checked again right before execution, so removed members stop acting
Model proposesa tool callThree checkstool · rules · screeningWaits for a personawaitingApprovalSlack · #procurementWants to call erp.inventory.lookupRuns with Sam Lin's permissionsApproveDenyRuns and is recordedwho asked · who approved · as whomThe same card works in the consoleRead-only tools an owner marksas automatic can skip the wait

Approval where work happens:
tool calls pause for a person

By default, every tool call the model proposes waits for someone to approve it in the console or on a Slack card. Both use the same state machine. After approval the platform checks permission again before running, and a call that already ran is never run twice.

  • Business rules can require approval for certain tools, or forbid certain actions outright
  • When content screening escalates a turn, every tool call in it needs approval
  • Read-only tools an owner marks automatic can run unattended, and one switch turns that off
Workspace monthly budget62%70%90%Under 70%the model the member choseClaude Opus by default70% to 90%steps down to a cheaper modeland says so on the replyPast the ceilingrefused with a reasonnever borrows next monthLedger · one row per model call10:42 procurement · sam.lin · claude-opus · 3,120 tok · chat10:43 procurement · sam.lin · claude-opus · 1,840 tok · tool11:00 weekly-report · auto · claude-haiku · 22,400 tok · schedule11:05 kb-index · system · claude-haiku · 8,900 tok · index

Budgets that hold:
ceilings per workspace and per team

Every model call takes one path: check the budget, call the model, write the ledger. Near the ceiling, replies move to a cheaper model and tell the reader. Past it, requests are refused with a reason. If the member picked a model by name, the platform refuses the reply instead of silently switching models.

  • Monthly workspace budgets, rolling-window limits, and per-team ceilings
  • Each call records who, which assistant, which model, how many tokens, and why
  • The usage report answers "why was this reply so expensive"

Platform Security

What IT checks before go-live is already built in

Where data lives, who can sign in, and how to trace an incident are already set up when you roll the platform out.

Data and isolation

  • Runs on Google Cloud in Taiwan (asia-east1), with a tenant boundary per workspace
  • A workspace can have its own dedicated Cloud SQL database
  • Connector credentials sealed with AES-256-GCM, with short-lived grants per use
  • On-premises systems connect through a relay, so data does not have to move to the cloud first (compare cloud, hybrid and on-prem)

Identity and audit

  • Enterprise SSO with SAML and OIDC, with directory groups mapped to workspace roles
  • Delegated identity, tool approvals, and token budgets with step-down in every workspace
  • Tool approvals, membership changes, and credential use are recorded, and owners can export them
  • Slack, Microsoft Teams, and LINE share one identity, budget, and audit trail

Cloud and on-prem

Runs where you need it: cloud, hybrid or on-premises

The same assistants, integrations and governance. Whether your data may leave your network decides where the platform runs.

Cloud

Runs on Google Cloud asia-east1 (Taiwan), with every workspace its own tenant boundary. No machines to prepare.

For teams that want to start this week and can keep data in a Taiwan cloud region

Hybrid: platform in the cloud, data on your network

Your ERP, warehouse and databases behind the firewall connect through a relay you run. The relay dials out, so no inbound firewall rule is needed; database credentials never leave your network, and every query is filtered to the asking person's own rows.

For companies whose core systems live in their own data center

On-premises: the whole platform in your environment

Agent Lab runs in your data center or private cloud, so assistants, knowledge and audit records stay inside your network. We plan the environment and sizing with you.

For finance, healthcare and public-sector teams whose data cannot leave the network

Plan an on-prem or hybrid setup

Bring one agent scenario that needs governance

  • 45 minutes walking an approval and a budget through your own scenario
  • Need on-prem or hybrid? We'll look at your environment in the demo
  • Straight answers on data residency, permissions, and cost control
  • If it fits, a workspace for your team the same week

Want to try it first? Sign up and open a workspace, then book a demo when you're ready.

Book a demo

Optional, helps us schedule faster