Agent Serve · Workflows
Connect triggers, AI agents, code and business tools in the workflow builder. Follow every run, and add a person where decisions need one.
Agent Serve Workflows is the visual builder in Agent Serve, the AI workspace where teams build, deploy, and manage AI agents. Teams connect blocks for models, integrations, code, conditions and human approval, start runs manually, on a schedule, from a webhook or an integration trigger, and deploy workflows as an API, a chat page or a tool.
How it works
An Agent Serve workflow starts on an event, lets an AI agent do the work, checks the output and asks a person where it matters.
A schedule, a webhook, a new table row or an event in a connected app starts the workflow.
The Agent block uses the model, instructions and tools you chose, and decides which tool to call.
Guardrails test for valid JSON, a pattern, unsupported claims or personal data before the run moves on.
The Human block pauses the run and alerts the approver, who resumes it from a page.
In depth
Agent Serve gives the workflow builder blocks for models, logic, code, data and people, and starts them from the apps you already use.
Each block in the Agent Serve workflow builder shows what it does in a sentence. Connect outputs to inputs and test a run without leaving the builder.
Agent Serve workflows start by hand, on a schedule, from a webhook, from a table row, or from an event in a connected app.
Agent Serve speech-to-text and text-to-speech blocks turn call recordings into transcripts and written replies into audio.
Describe the process in Agent Serve Chat and it proposes the blocks. Each change shows Accept and Reject before it lands in the workflow.
By the numbers
300+
Blocks in the workflow builder
400+
Integration triggers that can start a run
20
Parallel branches in one run
3
Ways to deploy: API, chat page and tool
Counted from the Agent Serve block and trigger registries and rounded down.
Capabilities
An Agent Serve Agent block takes a model, instructions, tools, files and memory, and decides which tool to call.
Agent → CRM lookup → Draft reply
The Agent Serve Human block pauses a run until a person responds, with optional Slack or email alerts to approvers.
Draft → Review → Send
Deploy an Agent Serve workflow as an API endpoint, a hosted chat page, or a tool on a workspace MCP server, a standard way for AI apps to call tools.
API · Chat · MCP
Every Agent Serve deployment is a version with who deployed it and when. Load an earlier one or Promote to live.
Version 12 · Promote to live
Agent Serve Logs show each block’s inputs, outputs, model, tokens, cost and duration, and let you retry a run.
Trace · Tokens: 1,204 in · 318 out
Agent Serve Function blocks run JavaScript or Python, and Table, Knowledge and API blocks bring in your data.
Function · Table · Knowledge · API
Examples
Three processes financial services teams build in Agent Serve Workflows, each with a person at the decision point.
An Agent Serve schedule picks the customers due for review, an agent gathers their records and drafts the refresh note, and an analyst signs off.
An alert arrives by webhook, an Agent Serve agent summarizes the customer’s history from tables, Guardrails mask personal data, and an investigator decides.
A new claim starts an Agent Serve workflow that extracts the policy and loss details, checks cover, and routes simple claims to fast track.
Examples show what teams can build in Agent Serve. They describe patterns, not named customer deployments.
Under the hood
| Triggers | Start (manual, chat and API runs), Schedule, Webhook Trigger, table row events and 400+ integration triggers. |
|---|---|
| Schedules | Every 1 to 1,440 minutes, hourly, daily, weekly, monthly or custom cron, in a chosen time zone. |
| Loops | For Loop, For Each, While Loop and Do While Loop, 1,000 iterations by default. |
| Parallel | Parallel Count or Parallel Each, up to 20 branches. Nesting up to 10 levels. |
| Wait | Up to 5 minutes, or up to 30 days with Async. |
| Code | Function blocks run JavaScript or Python. |
| Deploy | API with curl, Python, JavaScript and TypeScript samples, streaming and async; a hosted chat page; or a tool on an MCP server. |
| Versions | Rename, Add description, Load deployment and Promote to live. |
| Run trace | Per block: model, provider, tokens, time to first token, cost, duration and retries. |
Before and after
| Work | Before | With Agent Serve |
|---|---|---|
| How the process is defined | BeforeA script one engineer understands, or steps in a procedure document. | With Agent ServeBlocks in the workflow builder that each say what they do. |
| Where a person decides | BeforeAn email thread, with the decision lost in someone’s inbox. | With Agent ServeA Human block that pauses the run and records the response. |
| Changing a live process | BeforeEdit in place and hope nothing breaks. | With Agent ServeDeploy a new version, and Promote to live an earlier one if needed. |
| Finding out what went wrong | BeforeAsk around and rerun it by hand. | With Agent ServeOpen the run, follow the trace block by block, and retry it. |
Governance
Agent Serve Workflows check outputs, keep secrets out of sight and run from a deployed version, with a record of each run.
Security and governanceThe Agent Serve Guardrails block checks for valid JSON, a regex match, unsupported claims against a knowledge base, or personal data, which it blocks or masks.
Agent Serve secrets are stored in settings and referenced by name, and Function blocks can be limited to selected secrets.
Agent Serve access policies let admins hide each deploy option, API, MCP or Chat, for a group.
An Agent Serve schedule turns itself off after 100 failures in a row, so a broken run does not repeat forever.
Workflows run from a deployed version, so a change is reviewed before it reaches production runs.
Every workflow run records each block’s inputs, outputs and errors, so a reviewer can trace how a result was reached.
A Guardrails block detects personal data and blocks or masks it before text reaches the next step.
Admin, write and read roles per workspace, plus permission groups that limit which models, blocks and integrations a team can use.
Use your organisation’s own model provider keys at organisation or workspace level.
Sign in through your identity provider with SAML 2.0 or OIDC, and provision people with SCIM.
Run Agent Serve in your own cloud or data centre, next to the systems it reads.
Integration triggers
Works with
Describe a workflow in Chat, accept its changes in the builder, and deploy it from the conversation.
Call connected apps as blocks, and start runs from their events.
Choose the model each Agent block uses, per step.
Read and write table rows, and start a run when a row is inserted, updated or deleted.
Search knowledge bases from a workflow, and check answers against them with Guardrails.
Further reading
Disputes that fail reconciliation tend to sit in a side queue until the chargeback window has closed. Here is an end-to-end setup that finds them early, ranks them by deadline and drafts the pack, with an analyst signing every filing.
Cards and payments operations · 12 min readPeriodic reviews pile up because every file starts from scratch, even when nothing has changed. Here is a setup that refreshes the data, drafts the file and routes it by risk, so analysts spend their time on the reviews that need judgement.
Retail and business banking · 12 min readSampled call reviews find a missing risk warning weeks after the client acted on the advice. Here is a setup that checks every recorded advice call within a day and puts the quote in front of a compliance reviewer.
Wealth and advice · 12 min readQuestions
Agent Serve workflows run manually, on a schedule, from webhooks, from table row events, or through supported integration triggers. Deployment options also let other applications call a workflow as an API or a tool.
Yes. The Agent Serve Human block pauses a run until a person responds, and can notify approvers in Slack or by email.
Agent Serve deploys a workflow as an API endpoint, a hosted chat page, or a tool on a workspace MCP server. Each deployment is a version you can roll back to.
Yes. The Agent Serve Function block runs JavaScript or Python, and the API block calls any HTTP API.
Agent Serve Logs keep a trace of every run, with each block’s inputs, outputs, tokens, cost and duration. You can retry a run or ask Chat about the run history.
Yes. The Agent Serve workflow builder supports real-time editing by several people, with live cursors and undo and redo.
Last updated
Enterprise access
Agent Serve is available to teams by invitation. Tell us what you want to build and we will set up your workspace with you.