Governed Agents

Agents that act —
and can prove they were allowed to.

FlowClick's agent isn't a chatbot bolted on the side. It has governed access to your data, flows, files, models, and configuration — and every action it takes runs under someone's permissions, inside a budget, and onto the record.

What the Agent Does

A full-featured assistant with hands

Talk to it in an app, call it from a flow, or let it run as a job. It can read and write records, trigger flows, generate documents, send messages, browse the Marketplace, and change your application — through the same permissioned tools your team uses.

  • 53 governed tools covering records, tables, flows, files, content, versions, marketplace, models, and proposals
  • Three entry paths — chat (embedded in any public app), an agent step inside a workflow, or an in-process job
  • Classify, enrich, summarize, route, generate — as one step among 84, with its output available to the next
  • Embeddable in customer-facing portals with voice in and out, scoped to the tables you allow
  • Knows the platform's best practices and capabilities, so what it builds follows the rules a human reviewer would apply
 Support Triage — Flow
🎫 Ticket #4821 receivedTrigger
🤖 Agent: "Billing" · priority High · draft replyagent step
📈 Churn risk 0.71 → escalatescore step
👤 Account manager approves replyhuman task
Governance

Governed means five specific things

Most "AI agents" have all-or-nothing access and no memory of what they did. FlowClick's agent is a principal in the platform's security model, and it is held to it on every single tool call.

1 · Runs as a person

Every tool executes with the invoking user's identity, group memberships, and token scopes. The agent can never do more than the human who started it — the same PermissionEvaluator that guards the API guards the agent.

2 · Effect budgets

Declare per-account caps on classified side effects — outbound email, SMS, push, payments, external calls. The executor decides allow, warn, escalate, or refuse before the step dispatches, and writes an audit row for each decision.

3 · Dry-run before live

Run any agent under a QA session and its side effects are captured, not fired — so you can test what it would do. Put a human task in front of irreversible actions; row protection and retention holds apply to the agent like anyone else.

4 · Every run is a record

Chat, job, or in-process, each invocation is a durable AgentRun — principal, provider, tools called, outcome — joined to the trace of the flow that started it and the audit rows it produced.

5 · Kill switch and caps

Cancel any running agent job, cooperatively and immediately. Per-account caps on concurrent jobs and tool calls per run bound the blast radius of a runaway loop, and provider failures fail closed.

… and human review before it ships

When the agent wants to change an application, connector, mapping, or flow, it files a proposal. The platform validates it against real evidence and a person approves it. The self-building loop →

Models & Routing

Four providers. One loop. Routed by tier.

The agent loop is provider-agnostic. Native wire adapters for Anthropic, OpenAI, Google Gemini, and xAI Grok plug into one execution loop, so a flow written today runs on whichever model you choose tomorrow.

  • Claude by default, with Gemini, GPT, and Grok selectable per application or per step
  • Tier-based model routing — fast, standard, and premium tiers map to concrete models per provider; simple work goes to inexpensive models automatically
  • Bring your own provider credentials — keys live in a secret-safe connection, never in a variable
  • Streaming and non-streaming responses translated to one shape, so tool calls behave identically across providers
  • Voice: browser speech or provider TTS/STT for embedded assistants
Anthropic Claudedefault
Google Gemininative adapter
OpenAI GPTnative adapter
xAI Groknative adapter
↓ tiers: fast · standard · premium
🧭 Router picks the cheapest model that can do the jobper step
In Practice

Where governed agents earn their keep

Support & operations

Classify inbound tickets and email, draft replies in the customer's language, update the record, and route to the right team — with a human task before anything goes out to a VIP.

Finance & back office

Reconcile invoices against orders, chase overdue accounts within an email budget, and raise exceptions as human tasks with the evidence attached.

Building the platform itself

Ask for a connector to a new vendor or a mapping between two schemas. The agent drafts it, the platform tests it, and you approve it — the integration builds itself.

See a governed agent in your own process

Bring a process with real side effects — emails, records, approvals. We'll run it with an agent inside and show you every decision on the record.

Start Free Trial Schedule a Demo