Crescendo Agent Lab · Enterprise AI agent platform
Whose identity it acts under, what it costs, who approved it — every step is on the record. Agent Lab is the governed agent workspace from Crescendo Lab.
erp.inventory.lookup{ "sku": "A-1200", "warehouse": "north" }
Before you sign
If someone in the air-conditioner team can read the refrigerator team's reports through an assistant, that is a security incident. One retail chain's IT team put it plainly: they will not maintain a second set of access rules for AI.
With usage-based AI, one misconfigured schedule turns into a surprise invoice. Buyers do not want another dashboard. They want a ceiling that actually stops the spend.
When an assistant changes an order, you need three answers: who asked, who approved, and whose identity it ran under. A platform that cannot answer will not pass an audit.
How it works
These three mechanisms are in the platform's code. Every workspace has them from day one.
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.
By default every tool call the model proposes waits for someone to approve it — in the console or on a Slack card, one state machine behind both. After approval the platform checks permission again before running, and a call that already ran is never run twice.
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. A reply where the member named a model is never silently swapped — it is refused instead.
Build assistants
No code required. Each assistant has its own prompt versions, knowledge, skills, and tool bindings, and passes evaluations before it ships.
Guided setup, prompt versions you can roll back, and evaluation runs. Only assistants that pass the publish gate reach members.
Documents are indexed in the background and the most relevant sections are retrieved at answer time — with the asker's permissions, and with sources shown.
Write down how a job is done once and attach it to many assistants. Import skill packs from GitHub, with versions and a view of who uses each one.
Connect your MCP servers or HTTP APIs, including OAuth. Credentials are sealed and each use gets a short-lived grant; you choose whose identity a tool acts as.
Weekly reports and daily checks run on a schedule. Each run opens an ordinary conversation, under the same budgets and approvals as everything else.
Write in plain language what an assistant must never do, must do first, or must hand to a person. Tool bans are enforced, not just suggested in a prompt.
Channels
The console is home. Slack and LINE are where colleagues spend the day. Every entry point shares one identity, budget, and audit trail.
Conversations, approvals, and settings, in English, Traditional Chinese, Simplified Chinese, Japanese, or Thai.
@mention the assistant in a channel and it answers; approval cards appear right in the thread. Admins decide which assistant serves which channel.
What a manager asks for in a LINE group becomes a task on a shared board, with a daily digest and CSV export.
Agent Lab is also an MCP server. After OAuth consent, clients like Claude call the tools your workspace exposes as the member who connected them — never with more access than that member has.
Each assistant has its own API key, so your systems can bring it into existing workflows under the same budgets and audit trail.
@mention the assistant in a Teams channel and it answers, with approval cards right in the conversation. As with Slack, admins decide which assistant serves which channel.
Platform
Inside the same tenant boundary, a few building blocks companies actually use.
Upload a folder and get an address. Static sites publish directly; projects with a Dockerfile build and run in an isolated cluster. See Deploy
A workspace can have its own PostgreSQL instance. The platform provisions it and manages its users; your code connects directly.
Every model call is one row, viewable by assistant, member, or source, so the monthly bill reconciles.
Tool approvals, membership changes, and credential use are recorded, and owners can export the whole trail for audit or security review.
Claude by default, with OpenAI-compatible models available. Whichever you choose, every call passes the same budget and metering gateway.
Facts worth keeping are extracted after a conversation and read before documents next time, labeled separately in the reply. Curate or delete them in the console.
Our first customer is us
Franky is the AI coworker Crescendo Lab's staff @mention in Slack, and every answer cites a source. The identity, permissions, budgets, approvals, and audit behind it are Agent Lab — now a multi-tenant platform any company can open a workspace on.
Security and deployment
Where data lives, who can sign in, and how to trace an incident are part of the platform, not settings added during rollout.
We will reach you at the email you left within one business day to set a time.
FAQ