What Are Nightshift Agents?
A plain-English overview of how Nightshift's AI agents work, what they do, and how they operate without you lifting a finger.
Nightshift agents are specialized AI workers that handle specific parts of your business — bookkeeping, operations, payroll, sales, and reporting. They run automatically on a schedule, check in every 6 minutes, and prepare work for your review. You approve the important stuff; they handle everything else.
The idea in 30 seconds
Imagine hiring five employees who work around the clock, never forget anything, and never need a day off. Each one is an expert in their area. They pull data from your systems, analyze it, take routine actions, and bring you the decisions that actually need your input.
That is what Nightshift agents do. They are not chatbots. They are workers.
The agent team
Tower — The boss
Tower is the command layer. It does not run your day-to-day operations — instead, it coordinates everything. Tower builds new features, deploys updates to agents, reviews incidents, and pushes improvements. Think of Tower as the home office that makes sure all the branch managers have what they need.
Sub-agents — Your specialists
Each sub-agent is trained for a specific role. Here is who does what:
| Agent | Role | What they do | Example tasks | |---|---|---|---| | Penny Banks | Bookkeeper | Accounting, reconciliation, transaction categorization | Prep invoice batches, flag discrepancies, reconcile deposits | | Dash | Operations Manager | Revenue recovery, estimate follow-up, job quality | Chase unsold estimates, collect payments, audit completed jobs | | Sterling | HR & Payroll | Payroll prep, compliance, employee management | Prep payroll for your approval, track filing deadlines, monitor benefits | | Monroe | Sales & Marketing | Pipeline tracking, campaign analysis, lead scoring | Track lead sources, analyze campaign ROI, monitor close rates | | Chase | Reporting & Fleet | KPI dashboards, fleet tracking, performance reports | Generate scorecards, track vehicle maintenance, build weekly reports |
Primary agents — The location managers
If you have multiple locations, each one gets a primary agent that acts as the general manager for that location. The primary agent runs on its own dedicated machine and orchestrates all the sub-agents for that business.
For example, your Lexington location has a primary agent called Knox. Knox runs Penny, Dash, Sterling, Monroe, and Chase specifically for Glass Doctor of Lexington. If you add a Houston location, it gets its own primary agent called Clutch doing the same thing with Houston data.
How agents work
The heartbeat cycle
Every 6 minutes, each agent checks in with Tower. This is called the heartbeat. During a check-in, the agent:
- Pulls the latest code and instructions from the central repository
- Checks for new directives — tasks that Tower or you have assigned
- Reports its current status (healthy, busy, or in trouble)
- Picks up any config changes or improvements
You can see every heartbeat in the Command Center's Heartbeat tab — a live pulse showing that your agents are alive and working.
Scheduled tasks
Agents do not just sit around waiting for you. They have schedules:
| Time | What happens | |---|---| | 5:45 AM | Morning reports generated for each location | | 7:00 AM | Daily accounting run — Penny categorizes transactions and preps batches | | Every 2 hours | Dash follows up on unsold estimates | | 9:00 AM | Dash starts payment collection on outstanding invoices | | 6:00 PM | End-of-day reports generated | | 7:00 AM Saturday | Weekly summary reports | | Wednesday 8:00 AM | Sterling preps payroll for your review |
All of this happens without you doing anything. When you open the dashboard in the morning, the work is already waiting for you.
The approval model
Here is the critical part: agents prepare, you decide.
By default, no agent takes any action that affects your finances, customers, or data without your explicit approval. When an agent wants to do something important — like submit payroll, send a follow-up email, or close out a batch of invoices — it creates an approval request that appears in your Command Center.
You review it, check the numbers, and click Approve or Reject. The agent then executes (or backs off).
Over time, you can promote routine tasks to run automatically. For example, once you trust that Penny categorizes transactions correctly 99% of the time, you can promote that task to autonomous mode — she will still do it, but she will not wait for your sign-off every time. This is controlled in the autonomy settings in Command Center > Admin.
Autonomy levels
Each agent task has one of four autonomy levels:
| Level | Name | What it means | |---|---|---| | 0 | Report Only | Agent identifies and recommends but never acts. You must explicitly approve. | | 1 | Act & Notify | Agent executes the action and notifies you what it did. You review after the fact. | | 2 | Autonomous | Agent handles it entirely. You see it in the logs but no approval needed. | | 3 | Silent | Fully autonomous, minimal logging. Reserved for low-risk routine tasks. |
Most tasks start at Level 0. You promote them as you build trust. Financial tasks (payroll, invoicing, payments) stay at Level 0 by default and require your explicit sign-off every time.
What happens when something goes wrong
Agents are built to handle problems on their own first. When an agent hits an issue:
- Check the playbook — a shared database of known fixes. If the problem has been solved before, the agent applies the fix automatically.
- Self-heal — if a connection drops or an API errors out, the agent retries with backoff.
- Escalate — if it cannot fix the issue, it files an incident in the Command Center. You see a yellow or red alert, and Tower reviews the problem to push a permanent fix.
You can see all incidents, self-healing actions, and escalations in the Command Center.
Agents learn over time
After every task, agents capture what they learned — patterns in your data, API quirks, better ways to do things. These learnings are shared across all agents so the whole system gets smarter.
For example, if Penny discovers that a certain vendor always sends duplicate invoices, she logs that learning. Next time, she catches it automatically. If Dash notices that estimates over $10,000 take longer to close, he adjusts his follow-up timing.
You do not need to train your agents manually. They improve through experience.
Real examples of what agents do
To make this concrete, here are real tasks your agents handle every day:
Penny Banks — Tuesday morning: Penny pulls all deposits that hit the bank account yesterday. She matches each deposit to the corresponding ServiceTitan invoices. If a deposit does not match (maybe a customer paid a different amount than invoiced), she flags it for your review. She then categorizes the transactions in QuickBooks and prepares a reconciliation summary. Total time for you: zero. You see the summary in your Brief.
Dash — Every 2 hours: Dash checks for estimates that were sent but never sold. If an estimate is 3 days old with no response, Dash drafts a follow-up message and creates an approval request. If you approve, the message goes out. Dash also checks completed jobs for quality issues — missing photos, incomplete forms, or unusually low ticket amounts that might indicate a missed upsell.
Sterling — Wednesday morning: Sterling pulls hours and rates from Gusto, cross-references with ServiceTitan timesheet data, and builds the payroll summary. He flags any discrepancies (like a tech who clocked 50 hours in Gusto but only has 42 hours of dispatched time in ServiceTitan). You get a clean approval package with everything you need to review.
The self-healing system
One of the most powerful parts of the agent system is self-healing. When something goes wrong — an API connection drops, a data format changes, a vendor does something unexpected — agents do not just fail and wait for help.
They follow a playbook — a shared database of known problems and proven fixes. If Penny has seen a QuickBooks sync error before and knows the fix, she applies it automatically. If the fix works, she logs it. If it does not, she escalates.
Over time, the playbook grows. Problems that used to require your attention get handled automatically. The system gets more reliable every week.