This article is part of our AI Agents for Business hub. Want help applying it? Talk to eSEOspace.
Choose managed if you have no one to run the agent day to day. Choose in-house if the agent is core to your business and you can staff it. You can also split the work, and the split matters more than the label.
"Managed" also means two different things. It can mean a vendor hosts the software, or it can mean an outside team operates the agent for you. This guide treats those as separate decisions.
It is part of our AI agents for business hub and was checked against primary sources on 9 November 2026.
eSEOspace is a member of the Claude Partner Network, and Anthropic products appear below as examples.
Key Takeaways
- Hosting an agent and operating an agent are separate jobs, and a vendor may do only the first.
- An agent needs ongoing testing, and Anthropic pairs automated evals with production monitoring and human review.
- We found no reliable data on what managed or in-house operation costs, so we give none.
- Platform risk is real: OpenAI scheduled Agent Builder to shut down on 30 November 2026.
- Whichever model you pick, keep ownership of your test cases, logs, instructions and data.
The three operating models
| Model | Who hosts it | Who changes and tests it | Who you call when it fails |
|---|---|---|---|
| Hosted platform, your team operates | Vendor | Your team | Your team, then vendor support |
| Outside team operates | Vendor or the outside team | The outside team | The outside team |
| In-house | You | Your team | Your team |
A hosted product such as Claude Managed Agents falls in the first row. Anthropic's overview describes managed infrastructure and a built-in agent loop. It still leaves you to decide the agent's tools, instructions and permissions.
That is the point to check before you buy. Paying for hosting does not buy you someone who notices the agent got worse.
What "keeping it running" actually involves
Agents change without anyone editing them. Models get updated, connected software changes, and your own policies move.
Here is the recurring work. Assign a name to every row, whatever model you choose.
| Job | What it means |
|---|---|
| Test suite | Keep real tasks and expected results, and run them after every change |
| Monitoring | Watch what the agent did in production |
| Transcript review | Have a person read samples of real conversations |
| Prompt and tool changes | Update instructions and connections, with sign-off |
| Access review | Check keys and permissions against what it needs now |
| Cost watch | Track usage against budget and set limits |
| Incidents | Stop it, investigate and tell the right people |
| Vendor changes | Read deprecation notices and plan moves |
On testing, Anthropic's guide to agent evals says: "In reality, 20-50 simple tasks drawn from real failures is a great start." Anthropic sells AI models, so read it as vendor guidance.
The same guide says: "Production monitoring kicks in post-launch to detect distribution drift and unanticipated real-world failures." Its point is that no single method catches everything, so it pairs evals with monitoring and human review.
How much of this do teams actually do?
The best survey we found is LangChain's State of Agent Engineering, run from 18 November to 2 December 2025 with 1,340 responses. LangChain sells agent tools, and its respondents answered a public survey, from engineers and product managers to business leaders and executives. Technology made up 63% of respondents, so it does not describe a typical small business.
| Practice | Share of respondents |
|---|---|
| Some form of observability | 89% |
| Detailed tracing of individual steps | 62% |
| Offline evaluations on test sets | 52.4% |
| Online evaluations | 37.3% |
Quality was the top barrier to production, cited by about one-third of respondents. Even in this technical group, many teams had not set up evals.
If that is true of this group, ask who would do this work at your company.
When managed makes sense
- You have no one who can read logs or test an agent.
- The agent supports the business but is not what makes you different.
- You would rather pay for upkeep than hire for it.
- You are running one or two agents, not a portfolio.
Ask any outside provider what they will actually do each month. A list of duties, a response time and a report you can read are worth more than the word "managed."
When in-house makes sense
- The agent touches the core of how you make money.
- Your data or contracts limit who can handle it.
- You already have engineers with spare capacity.
- You plan to run many agents, so the skills will keep paying off.
In-house has a staffing risk that is easy to overlook. If one person built it and leaves, you have the same problem as a vendor disappearing.
A split to consider
You do not have to choose one side for everything. A sensible division keeps decisions with you and hands the routine work to whoever is best equipped.
| Keep in-house | Can be handed off |
|---|---|
| Deciding what the agent may do | Building and updating connections |
| Approving high-impact actions | Running the test suite |
| Owning the test cases and results | Watching dashboards and alerts |
| Owning data, logs and instructions | Applying model and platform updates |
OWASP's AI Agent Security Cheat Sheet says: "Require explicit approval for high-impact or irreversible actions." That approval should stay with someone who answers for the result.
Plan the exit before you sign
Vendors change direction. OpenAI's deprecations page lists a 3 June 2026 notice that Agent Builder would shut down on 30 November 2026. The Evals dashboard and API have the same date, so a team could lose its builder and its test tooling together.
Ask for these in writing, whichever model you choose:
- Export of instructions, tool definitions and configuration
- Export of test cases and results
- Export of logs and conversation history
- Documentation good enough for a new team to take over
- A named notice period for pricing and service changes
Our AI agent vendor questionnaire puts these into a form you can send. For the buy-or-build decision itself, see build vs buy an AI agent.
Next step
Write the eight jobs above on one page and put a name next to each. If a row stays blank, that is your answer about what you need to buy or hire. Our AI agent development page says: "We can look after it on a monthly plan, or hand it over to your team with documentation."
Frequently asked questions
What is the difference between a managed AI agent and a hosted one?
Hosted means someone else runs the infrastructure. Managed, in the operating sense, means someone also tests, updates and monitors the agent. Ask which one a vendor means.
Is managed cheaper than in-house?
We found no reliable public data to answer this, so we will not guess. Compare a written monthly scope against the actual time and people the work would take at your company.
How often should an agent be tested?
Anthropic's guide describes running regression tests continuously, and pairing them with production monitoring. A sensible minimum is after every change to the model, instructions, tools or data.
What should I own no matter who runs the agent?
Own the instructions, test cases, logs, data, and the authority to approve risky actions. Those are what let you change providers or stop the agent.
Can a managed provider take responsibility for what the agent does?
Only as far as your contract says. Have your attorney read how the agreement handles errors and liability, because it will decide who pays when something goes wrong.
