Skip to main content
VantaSoft
Back to Insights

AI Agent Maintenance: What Keeping Our 20 Agents Running Takes

AI agent maintenance is updates, sign-ins, rule changes and catching quiet failures. What 30 days of it took for our 20 agents, and what stays with you.

Rollout and oversight 10 min read

AI agent maintenance is the work that keeps an agent doing its job after launch. Someone installs software updates without breaking anything, keeps the agent's sign-ins to your tools and to its AI model working, rewrites its rules when your business changes, and notices when a scheduled job stops or quietly goes wrong. That work comes up every week.

VantaSoft runs on 20 AI agents, and we keep them running ourselves. We went back through our own records for the 30 days from October 18 to November 16, 2026, and sorted the upkeep by kind.

Kind of upkeepWhat happened in our 30 daysWho handled it
Software updatesHermes Agent, the free software our agents run on, put out 8 regular releases. Each one goes to a test agent first.Our engineering agent
Rule changesAn owner set or changed a standing rule for at least one agent on 29 of the 30 days.The owners decided, and an agent wrote each rule down with a name and date
Jobs changing hands9 recurring jobs moved to a different agent, and we added 7 agents.The owners decided, and the agents moved the work
Shared limitsA morning spreadsheet check hit Google's limit on reads per minute on October 19 and 23.We moved all but one job off the hour and half hour
HousekeepingOur test server filled to 96 percent with old copies of updates.Our engineering agent cleared 38 and fixed the cleanup
Jobs collidingOn November 7, two daily jobs saved one file in the same second and broke it.Our content agent made each job wait its turn

Underneath all of that, 174 scheduled jobs now run across 17 of our agents. On October 23 there were 104.

What does AI agent maintenance include?

AI agent maintenance covers four kinds of work. The agent's software and AI model get updates. Its sign-ins lapse or get checked. Its rules change when your business does. And someone has to catch jobs that stop, slow down or turn out wrong work without any error message.

Updates to the software and the AI model

Hermes Agent put out a regular release about twice a week in our 30 days. Some fix problems and some change how the agent behaves, which is why each update reaches a test agent before any working one.

On October 11, just before our window, a change moved where our agents looked for the small program they use to reach their AI model. Every agent started failing with the same message, that its version did not support the model. We patched it that afternoon and fixed it for good in a new release the same day. The agents' work had not changed at all, only one setting underneath them.

AI companies also retire models on a published schedule, and moving to the next model means retesting every job. We cover those schedules in build your own AI agents or buy a managed service.

Sign-ins that lapse

An agent works through logins to your email, calendar, files and the AI model itself, and logins expire. On the agents we run for customers, a Claude sign-in has to be renewed every month. The agent messages the person who signed it in three days before it lapses, and again if it does. Only that person can renew it.

Some checks can only be passed by the account holder, such as a code sent to their phone, so no provider can do that part for you.

Rules that change with the business

Rule changes came up more often than anything else. On 29 of our 30 days, one of our three owners gave an agent a new standing rule or changed an old one. A few examples from the month:

  • Every task an agent adds to our CEO's to-do list now needs a due date.
  • Our content agent switched to a newer image model on November 5.
  • On November 15, five recurring jobs moved from the CEO's assistant agent to the new assistants for our other two owners, and a sixth went to our finance agent.

Each one had to be written into the right agent's instructions, and any agent that relied on the old rule had to be told. We explain how we record corrections so they stick in how to train an AI agent.

Jobs that stop or go quietly wrong

The spreadsheet check that failed in October did at least fail loudly. Google allows each connected app 300 spreadsheet reads a minute, and the app we connect through is shared with that service's other customers. With 57 of our 104 jobs starting at the top of the hour, 24 more on the half hour, and that service's other customers doing the same, the reads ran out. A rerun a few minutes later worked each time, so we spread the start times instead.

The full test server and the colliding jobs were quieter. Neither showed an error until the day something stopped working. When we caught the server at 91 percent, about six more updates would have filled it, and then no new update could have been tested at all. It reached 96 percent before the cleanup was fixed and 38 old updates were cleared, which brought it down to 38 percent. The file the two jobs broke would not open until it was repaired by hand.

Quiet failures are the ones to watch, because a job that did not run looks the same as a quiet day. We wrote up five more in what went wrong when we gave AI agents real work, including one where six jobs went silent for an hour.

How often does an AI agent need maintenance?

Plan on a quick check every day and some upkeep every week. In our 30 days, the agent software had a new release about twice a week. An owner changed some agent's rules on all but one day. And the scheduled jobs we watch grew from 104 to 174.

One agent doing one job gets the same kinds of upkeep in smaller amounts. Software updates arrive at the same pace no matter how many agents you run. The rules change whenever your business does: a new price or a new approver. Give the daily check to one named person, since it is the easiest to drop.

What stays with you as the owner?

Even with a provider doing the upkeep, four things stay with you, because only you know them or can do them.

  • Telling the agent when something changes. It cannot know about a new policy until someone writes it down.
  • Approving the actions it is set up to ask about, such as payments and messages to customers. See human in the loop for AI agents for which ones to keep.
  • Renewing sign-ins when the agent asks, and answering codes sent to your phone.
  • Reading a few finished results each month. The seven-question monthly review takes about twenty minutes.

An AI agent maintenance schedule you can copy

Use this to plan the upkeep for one agent, whether a provider runs it or your own staff do.

How oftenWhat gets doneWith a managed serviceIf you run it yourself
Every dayConfirm every scheduled job ran, and read any failureThe providerYou or someone on staff
About twice a weekInstall software updates after trying them on a test agentThe providerYou
Every weekCheck that storage is not filling up, and clear old copiesThe providerYou
When your business changesUpdate the agent's rules and who it answers toYou say what changed. The provider edits the instructions.You
When people join or leaveChange what the agent can reachYou decide. The provider makes the change.You
Every monthRenew the AI sign-in when asked, and read three finished results at randomYouYou
When an AI model is retiredMove to the new model and retest every jobThe providerYou

If most of the right-hand column would land on one busy person, that is the real cost of running it yourself. What Hermes Agent costs a small business to run puts staff time on that bill line by line.

What does AI agent maintenance cost?

It depends on who does it. SoftTeco, a software development firm, estimates that maintaining and scaling a custom-built AI agent runs from $5,000 to over $50,000 a year. With a managed service, a monthly fee covers the upkeep. If you do it yourself, the cost is your team's time.

Our own hosting and maintenance starts at $475 a month. It covers your agents on their own private server, monitoring, updates and maintenance, email support during Pacific business hours, and small changes as your work evolves. The build starts at $2,500 one time and includes 16 implementation hours. For every cost line side by side, see how much an AI agent costs.

Frequently asked questions

What does post go-live support look like for a custom AI agent?

Ask for it in writing. Good support covers watching every scheduled job, installing and testing updates, fixing sign-ins and connections that break, and making small rule changes as your work changes. Ask how you will hear about a job that did not run, and how fast someone replies. Ours is email support during Pacific business hours.

Does an AI agent need to be retrained?

Usually not in the sense people mean. Agents built on models like ChatGPT or Claude are not retrained on your data. When the job changes, you update the agent's written instructions and reference pages. That updating is ongoing, and on our team it happened on 29 of 30 days.

What happens if nobody maintains an AI agent?

It keeps running on old rules and old logins until something breaks. Some breaks are loud, like an error on a failed sign-in. Others are silent, like a job that stops running or a report built on last quarter's prices. The silent ones can go weeks before anyone notices.

Can I maintain an AI agent myself?

Yes, if someone on your team can own it. That person checks the jobs every day, installs and tests updates as they come out, fixes whatever breaks and writes down every rule change. Put it in their week by name, with time set aside.

Where to start

Copy the schedule above and put a name next to every row. Any row without a name is upkeep nobody is doing yet. To see what our agents do between those checks, read our AI agent examples.

If you would rather have an agent built around your workflows and kept running for you, Start Automating. It starts with 30 minutes on your workflows.

Five quick questions about where your team's time goes. You'll get the one agent we'd start with, and what it would take off your plate.

Take the quiz

Put your first agentto work.

We build custom AI agents around your processes and the tools you already use. It starts with 30 minutes on your workflows.