Awake
The moment the agents are switched on. Each one takes its post inside a real workflow — with an owner, a role, its skills, its tools and its channels — and the company starts working with two kinds of crew.
Awake is the moment the agents designed in Align are switched on. An agent is a configuration, not code: an owner, a prompt, a model, tools, skills, channels — and the moments it wakes. From then on the harness wakes it by itself: on a schedule, when the business moves, or when someone writes.
Agents should show up like crew, not like software. Every agent gets one harness and its own uniform: the same rules, the same map of the company and the same gate on every tool — plus its own role, owner, leash and channels. People stay at the decision points.
An owner, a prompt, a model, a look. The agent is a configuration, not code — an edit reaches its very next turn. Only an active agent runs.
In training it shows what it would do and waits for your OK. Autonomous, it acts and reports. Manual, it works only when asked. Training is a phase, not a destination.
Team chat, web chat, mailboxes, meetings, voice, WhatsApp, SMS — plus schedules, business events and other agents. When in doubt, it drafts instead of sending.
Every five minutes a dispatcher wakes what’s due. Schedules run versioned skills, breakers switch off a failing schedule and tell its owner, and cut-off work resumes from a checkpoint.
One gate on every tool call. Proof before claims. Escalations climb timed levels — and money, policy, safety, legal and access always reach a person.
Every run becomes a card: a sentence of what it did to the business, and a verdict on a simple question — would the business miss it?
Five moments define the agent: who it is, what it does, where it lives, when it wakes, and how much leash it gets.
Tools, skills, channels and peer agents are assigned, not coded. What it may do depends on who is writing to it.
In training, the agent proposes and waits for an OK. A certification run can check that its tools are really wired and that it doesn’t invent tool calls.
Status goes to active — the only status that runs. Schedules and business events start waking it; a paused agent never fires.
What it can’t or shouldn’t decide becomes an escalation: a card in the department chat, then a message to the lead. A decision wakes the agent again.
Read the run cards. Edit a skill and the next run follows the new page. A lesson learned by two agents can travel to the whole fleet.
Answers in chat, mail, WhatsApp or voice — with the permissions of whoever is writing. Guests are offered a human.
Wakes on the clock and runs a versioned skill. Breakers stop a schedule that keeps failing and tell its owner.
Wakes when the business moves — a record changes, a document is signed, a field update lands. Each event is claimed once.
Sweeps for what needs attention and speaks only when something is actionable. Nearly half of all runs are this shift.
Asks other agents for help, up to three hops deep, and hands decisions to people through the escalation ladder.
Built from the tool calls, with no model: died, failed, advanced, wrote, consulted or untouched. A business verb counts only with mechanical proof.
Would the business miss it? None, low, medium or high — scored from probabilities, not prose.
An agent may only say something happened if a tool proved it in the same turn. Unproven claims are rewritten before they reach a chat app.
Writes and payments for a customer recognized by phone need a one-time code — valid ten minutes, at most three an hour.
Ten straight configuration failures switch a schedule off and message its owner. Infrastructure failures back off instead.
Work that gets cut off continues from a checkpoint. In 30 days, 89% of 5,330 long runs completed.
Events are claimed by a dedup key and writes go through a ledger, so retries don’t repeat work. A test ran 500 events on 25 links exactly once each.
A schedule can declare what success means — because 17 of 247 “successful” runs once reported failure in their own text.
The same rules, the same map of the company and the same gate on every tool — for every agent. This is where Awake begins.
Each agent’s own owner, prompt, model, look, tools, skills, channels and leash. Configuration, not code.
Schedules, business events, queues, breakers and durable runs. The harness wakes the agent — nobody has to.
Versioned skills (1,528 of them, 10,709 saved versions), memories, and lessons shared across the fleet. The crew gets better between shifts.
One workflow, one owner, a few agents in training. Read every run card.
Schedules and business events switched on. Breakers and escalation ladders in place.
Agents meet people where they already work — chat, mail, voice, WhatsApp.
Wake on the business, not just the clock. Retire the runs nobody would miss.
A fleet that learns overnight, with people at every decision that matters.