Five Kinds of Agent Builders, Reviewed: Canvases, Chat Agents, Frameworks and More

There are now so many ways to build an agent that "which tool should I use?" has become nearly impossible to answer. So we asked a better question: which kind of tool?

Almost every product on the market has one of five basic designs. Brands come and go and pricing changes monthly, but the designs are stable, and each one behaves in its own way. Over three weeks I built the same three agents with a typical tool of each kind and kept notes on what hurt.

How we tested

We picked three jobs because readers keep asking about them:

1. The digest. Every Monday, check six competitor pricing pages and email a summary of changes.
2. The triage. Read incoming support emails, tag them by topic and urgency, and log them in a spreadsheet.
3. The onboarding. When a new customer signs a contract, create a project, invite them to a shared folder and draft a welcome email for approval.

For each tool I scored setup time, reliability across twenty runs, how easy it was to debug a failure, and how predictable the cost was. Scores are out of five and reflect our own testing, not a lab benchmark.

The scorecard

The five kinds, up close

No-code canvases

You drag boxes onto a canvas ("read email," "ask the model," "write to sheet") and wire them together. The digest took me 40 minutes. A canvas is wonderful for showing a colleague what the agent does, which matters more than people think.

The trouble starts at the edges. When the triage agent hit an oddly formatted email, the failure appeared as a red box with a vague error. Fixing it meant adding branches, and by the end my tidy canvas looked like a subway map.

Chat-native agents

These live inside a chat assistant. You write instructions, connect a few apps and talk to it. Setup was the fastest of all: the digest took twelve minutes, mostly spent on wording.

Reliability was the weakest. Across twenty runs of the triage job, the same email was tagged "billing" on some runs and "account access" on others. These are great for errands where a human glances at the result, and shaky for anything that feeds a spreadsheet other people rely on.

Automation platforms with AI steps

This category was the quiet winner. These are the long-established "when this happens, do that" tools, now with model-powered steps added. The model handles only the fuzzy parts, like tagging, summarising and drafting, while the rest is predictable plumbing that has been tested in production for years.

Onboarding, the job with the most moving parts, was most reliable here, with nineteen clean runs out of twenty. The twentieth failed loudly with a clear log, which is exactly how you want things to fail.

Code-first frameworks

If you write code, these libraries give you total control over your own tools, retries and tests. Debugging scored a five because I could read every step. Setup scored a two because I spent a full afternoon on authentication before the agent did anything interesting.

They're the right choice if the agent is your product. For a one-off internal errand, they're a power drill for hanging a postcard.

Managed agent platforms

These host your agents and bundle memory, permissions, monitoring and evaluation dashboards. The guardrails were genuinely good. By default, the onboarding agent couldn't email a customer without an approval step.

Cost was the hardest to predict. With usage-based pricing on top of model costs on top of a platform fee, I couldn't forecast my test month's bill until it arrived.

The verdict

• Start with an automation platform with AI steps if the process already exists and you just want the fuzzy bits handled.
• Use a chat-native agent for personal errands where you'll read the output anyway.
• Pick a canvas if the agent's main audience is a team that needs to understand it.
• Go code-first only if you're building something to sell, or something that must be tested like software.
• Choose a managed platform when governance matters more than cost: regulated work, big teams, many agents.

The most useful finding wasn't in the table. In every category, the agents that worked best were the ones doing the least. The kind of tool matters, but the size of the errand matters more.