Most comparison pages exist to make one option look inevitable. This one exists because we lose time — and so do you — when a project starts with the wrong shape. Two of the five options below need no vendor at all, and we have pointed people at them.
Read the table for the shape of the decision, then the section for whichever two options you are actually choosing between. Every claim here is about the category, not a scorecard of named products: tool capabilities change monthly, so verify current features yourself before you commit. The vendor scorecard is the question set we would want a buyer to bring to us.
The decision at a glance
| Option | Best when | Where it stops | Who runs it after |
|---|---|---|---|
| Build in-house | The agent is part of your product and you employ engineers who can own it | First production incident lands on a team with no agent experience | Your engineers, permanently |
| No-code automation Zapier, Make, n8n | The steps are deterministic: triggers, filters, field mapping | A step needs judgement on unstructured input | Your ops person |
| Hosted agent platform | Your use case matches what the platform already does | You need your own architecture, hosting, or audit trail | The vendor, while you subscribe |
| Generalist dev agency | The agent is one feature inside a larger app build | Agent-specific failure modes: drift, tool errors, silent wrong answers | You, after handover |
| Specialist agent studio what we do | Agents are the point, and someone has to run them once live | Small one-off automations — we are the expensive way to do those | Us, or your team after handover |
Build it in-house
This is the right call more often than vendors admit. If the agent touches your core product, the knowledge belongs inside the company, and outsourcing it means outsourcing the thing you compete on.
The cost is not the build — it is the second month. An agent that acts on real data needs someone watching output quality, handling tool and API failures, and deciding what happens when the model returns something confident and wrong. Teams that have not run one before usually under-plan that part. Choose in-house when you have named the engineer who will own it after launch.
No-code automation platforms
Zapier, Make, and n8n move data between systems on a trigger, with steps you define. When the work is deterministic — a form fills a CRM record, a paid invoice files a document — this is the simplest thing that works, and adding a model to it makes the system less predictable, not more capable.
The line is judgement. When a step means reading an inbound reply, deciding whether it qualifies, and picking the next action, deterministic branches stop scaling: every new case is another branch. That is the point where an agent starts to pay for its complexity. Plenty of production systems keep the automation platform for triggers and hand only the judgement step to an agent — n8n in particular is often the right co-pilot to an agent rather than a replacement for one.
Hosted AI agent platforms
A hosted platform is the fastest route to something running: the infrastructure, model access, and monitoring are already built. If your use case sits inside what the platform does, use it — a bespoke build of a solved problem is waste.
Two things to check before committing. First, where your data goes and what the vendor may do with it, which is a procurement question, not a technical one. Second, what leaving looks like: if the workflow logic lives in the vendor's console, your exit cost is a rebuild. Both are answerable in writing before you sign.
A generalist dev agency or freelancer
Sensible when the agent is one feature inside a larger application build, and the same team is doing the app. You get one contract and one point of contact.
The gap is the failure modes. Agent systems fail differently from web apps: they are confidently wrong rather than broken, they drift as prompts and models change, and they need evaluation harnesses and guardrails that a general web team has usually not built before. Ask for a system they run — not a demo — and how they detect a bad answer in production.
A specialist agent studio
This is us, so treat it as the pitch: we build on Hermes, OpenClaw, and custom stacks, we run agents in production for our own agency, and every case study on this site ends with what it does not prove. What you are buying is the part after launch — the governance, the failure detection, the operating discipline.
When we are the wrong call
Four situations where we will say no, or point you elsewhere:
- The workflow is deterministic. If nothing in it needs judgement, you want an automation platform, not an agent, and not us.
- It is one small task. A single scheduled script does not justify a scoping engagement.
- The process is not written down yet. An agent automates a process; it does not invent one. If nobody can describe the steps a human takes today, that is the work to do first — and it is free to do it internally.
- You need the capability in-house. If this is core product knowledge, hire for it. We will happily scope the blueprint your own team then builds.
What to ask any vendor, including us
Same questions, every vendor, answers written down:
- Who owns the code and the data, in writing, and where is it hosted?
- What happens to the system if we stop paying you?
- How does the agent detect that it has done something wrong, and who is told?
- Which production system of your own can you show us, with real numbers?
- What does month two look like — who watches it, and under what agreement?
The vendor scorecard turns that into a scored checklist you can run across every quote, and theproduction readiness diagnostic covers the operational side. Both are free and neither asks for an email before showing the result.