Released 2026-05-26. Consolidates seven releases that shipped together. The Agents module lands in five phases and the AI assistant gets persistent conversations.

Agents (preview) — the full module

Create AI agents that complete a single task end-to-end. The runtime loops the AI model + tool calls until it produces a final answer.

Four agent types — interactive (public chat widget), webhook (signed inbound POSTs), schedule (cron), playground (bound to a playground submission).
Task description + AI model picker — choose provider + model + max tokens. Optional team AI connection per agent.
Allowed tools — pick which platform-tool actions, AI tools, MCP servers, and workflows the agent may invoke. Without an explicit allow-listing, the agent can't call a tool.
Workflow triggers — list workflows the agent may trigger and describe when each should fire; the LLM decides at runtime.
Support files + AI calibration — upload PDFs / HTML / Markdown / JSON / CSV; click Calibrate to have the AI summarise them for the task. Review, edit, and approve before runs use the brief.
Approval rules — describe conditions in plain English ("pause before refunding more than $200"). The runtime pauses the run and emails approvers; one click resumes or rejects the run with audit history.
Scheduled agents run on cron with per-run history capped by the plan's agent_run_history_months.
Webhook agents accept external HTTP triggers; signature verified with the configured platform's standard header.
MCP server tools — agents may bind a platform tool in MCP mode and call its remote tools at runtime alongside platform actions.
Dashboard test-run + per-run timeline — kick off a test from the editor and watch the step-by-step LLM / tool / approval trace on the run detail page.

Interactive chat agents

Embeddable chat widget with a branded launcher, persistent multi-turn message history, and a configurable footer. The "Powered by Flowfn" footer is plan-gated by allow_remove_flowfn_branding (the same white-label gate used for playgrounds, forms, and visualizers).
Dashboard chat-test panel for owners to converse with their agent without leaving the editor.
Public chat URL for visitors who land directly on the agent (added in this rollup; refined in v1.3.0).

Playground ↔ Agent binding

Playgrounds carry an optional agent_binding pointing at an interactive agent. Save is plan-gated by allow_playground_agents.

Embed cheatsheet

The agent editor's Embed & branding tab has a Show embed snippets button that opens a side drawer with copy-paste HTML for the bootstrap script + raw iframe + an allowed-domains note.

Self-service data export — full agents coverage

The DSR JSON now includes agents (owned by the requesting user), agent_support_files (metadata only — binaries stay on S3 under access control), agent_runs, and agent_approvals. Export format bumped to 1.2. Embed tokens, webhook secrets, and bearer credentials remain excluded — they're credential material.

AI assistance

Persistent AI conversations with a side-panel history. Delete a conversation with a 60-day grace window before it's permanently removed.
Cross-tab resume — pick up a conversation on a different tab while a turn is still mid-flight.
AI Assistance polish — chat turns that produce a workflow / playground / form / visualizer show as that type on the requests page, playground-behavior rows are visible there, and AI-create dialogs start with an empty slug.
Data export trim + broaden — credit ledger entries are no longer in the self-service export (request via DSR if needed). The export now includes privacy / processing restrictions and AI preferences.

Legal & marketing

AI Disclosure §2 lists the Agents module explicitly. Cookie Policy §8a describes the iframe's no-cookie session model. READMEs in flowfn-app and flowfn-engine have new sections describing Agents and the agent API surface.

v1.2.0