Today we're shipping Tapotik Agents to every Pro and Enterprise workspace. Agents plan, execute and verify multi-step work across your tools — researching a refund, triaging a bug, reconciling an invoice — and they do it with a property most autonomous systems skip: they know when to stop and ask a human.
We've watched teams wire chat models into workflows for two years. The pattern is always the same: the model is smart enough to do the work, but the glue around it — retries, state, tool wiring, escalation — swallows weeks of engineering. Agents productize that glue.
Plan, execute, verify
Every run moves through three phases. First the agent plans: it decomposes your goal into steps and declares which tools it intends to use, so you can inspect intent before anything happens. Then it executes, chaining up to 25 tool calls with parallel dispatch where steps are independent — runs are 2.3× faster than our beta because of this alone. Finally it verifies: a separate pass checks the outcome against the original goal, and failed verification sends the agent back to planning rather than out the door with a wrong answer.
The whole run is visible as a live execution graph in the dashboard. Click any node and you see the exact tool call, its arguments, the response and the tokens spent. Nothing is a black box, because nothing invisible ever earns trust.
Approval gates, built in
Some actions should never be fully autonomous. Mark any tool with requires_approval and the run pauses before calling it — the agent presents what it wants to do and why, and a human approves or rejects from the dashboard, Slack or email.
- Scoped gates: gate a single tool, a category (anything that writes), or every action above a credit threshold.
- Context attached:approvers see the plan, the evidence gathered so far and the exact payload — not just an "approve?" button.
- Full audit trail: every approval, rejection and auto-resumed run lands in the workspace audit log with actor and timestamp.
Autonomy isn't the absence of humans. It's knowing precisely when a human matters — and pausing there, every time.
What beta teams shipped
Northwind's support team runs a refund-triage agent that verifies order state in Stripe, checks policy and drafts the resolution — humans approve the refund itself. Their handle time dropped 62%. Vectorline pointed an agent at their release process: it compiles changelogs from merged PRs, drafts the announcement and schedules the send, saving an engineer most of a day per release.
The common thread: agents win on work that is procedural but not predictable — too variable to hard-code, too repetitive for a person to enjoy.
Available today
Agents are live for Pro and Enterprise plans. Runs bill underlying model tokens plus 2 credits for orchestration, and every workspace gets 25 free runs this month to experiment. Start in the dashboard with a template, or go straight to the API — the docs cover the full run lifecycle, webhooks and approval endpoints.
We think this is the biggest step yet toward AI that ships work, not suggestions. We can't wait to see what you hand off first.