Always on. Rarely running.
Waveguide isn't a chatbot waiting for you to type, and it isn't a VM burning money around the clock. Your agent sleeps until something real happens, wakes, does the work, reports, and goes back to sleep. That one design choice is why it's affordable — and why nothing happens without a signal you can trace.
What wakes the agent
Three kinds of signals, each with a purpose. Between them, the agent's machine is suspended — state snapshotted, compute off.
Push events
A lead form fires, a reply lands in Gmail, a calendar invite changes, a campaign webhook arrives — the agent wakes within seconds, does exactly the work the signal calls for, and goes back to sleep.
Scheduled digests
Two or three times a day (by plan), the agent wakes to review everything, write your brief, and queue the decisions that need you.
Heartbeats
For sources that don't push, a lightweight check a few times a day — run by the cheapest model — catches anything that slipped through.
A day in the life
Not hypothetical — this is the loop, hour by hour.
- 8:04 AMMorning brief
Overnight summary in Slack: 3 campaigns healthy, 2 hot leads answered, 1 budget decision waiting for you.
- 10:12 AMLead arrives
Form fill wakes the agent. Enriched, scored 87, hot — escalated to you with a drafted reply referencing their portfolio size.
- 11:30 AMYou approve
One tap in Slack. The reply sends from your own Gmail. The lead, the score, and the touch are all in your ledger.
- 2:30 PMBudget decision
Meta CPL dropped 22%; the agent proposes raising the retargeting set $60 → $140/day. Within your cap, so one tap approves it.
- 4:45 PMMission progress
The outreach mission booked its seventh demo. Follow-up queued; the funnel card in your dashboard ticks forward.
- 7:00 PMEvening close-out
What changed, what it did, what's queued for tomorrow. Then the agent's machine suspends — you stop paying for compute until the next signal.
The shape of the system
Three layers, one guarded door. This is the architecture that makes the security promises checkable rather than aspirational.
Writes and runs code. Default-deny network. No credentials inside — ever.
Risk tiers checked, approvals verified, credentials injected, everything logged.
Reached only through the proxy, on allowed paths, within your caps.
Your agent, in a sandbox
Each tenant's agent runs in its own Firecracker micro-VM with default-deny networking. It writes and runs real code — but the only route out is the guarded door.
The egress proxy
Every authenticated call exits through one proxy that checks the action's risk tier, verifies approvals, and injects credentials after the request leaves the model. The model composes requests; it never sees a token.
Your accounts
Gmail, Calendar, Meta, Stripe, your MCP tools — reached only through the proxy, only on the paths you've allowed, only within your caps. Every call is logged on the way through.
The right model for the job
A single-model agent is either too dumb or too expensive. Waveguide routes every piece of work to the cheapest model that can do it well — that's the economics behind the flat monthly price.
Fast + inexpensive
Reads everything: digests, heartbeats, triage. Sorts signals into act / notify / ignore so the expensive model is never woken for noise.
Flagship
The default worker: campaign decisions, lead replies, creative direction, mission planning and execution.
Heavyweight
Escalation for genuinely hard problems (Growth and Scale plans) — gnarly attribution questions, complex mission re-planning.
Every model call is metered to the cent and visible on your Billing page, with automatic guardrails at 80%, 100%, and 120% of your included budget.
See the loop run on your own accounts
Five minutes of setup, and the first brief is usually minutes behind it.
Card required · nothing charged until day 15 · cancel in one click