Teams also keep a small set of recorded test cases and rerun them whenever the prompt or model changes, the same discipline as regression tests. Other visual builders, n8n and Zapier among them, now offer agent steps too, and the same rule holds there: the safety comes from what surrounds the step, not from the product.  Tracing tools from the observability world, whether hosted or open source, make the same point from the other side, since you can only constrain what you can replay. That is why a bounded step in an inspectable workflow is a defensible default rather than a timid one. ## What Activepieces AI agents cost and require You can self-host the MIT-licensed Community Edition for free, paid plans add hosted convenience and enterprise controls, and whichever provider you connect bills the AI model's tokens separately. ### Community Edition versus paid plans You can run, read and modify the code without paying the platform, because it is MIT-licensed. What you give up is the vendor's hosting and its paid-tier features. If you need single sign-on (SSO) beyond the Team tier, audit logs or embedding the builder in your own product, those go through the [sales form](https://www.activepieces.com/sales), so budget a conversation rather than a checkout. SOC 2 is in progress, not certified, so a procurement team that needs the report today can't have it yet. Run it yourself and the operations land on you: you provision the servers, apply upgrades, run backups and secure the instance. ### Model provider costs sit outside the platform fee Whichever plan you pick, the agent step calls a model you pay for separately. The figures below, from [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing), show how far apart one vendor's own tiers sit, per million tokens:  - Fable 5.1 runs $10 to $50, so a reasoning-heavy step on it can cost many times more than the same step on a smaller tier. Opus 5.5 runs $4 to $20, which suits long-running work where the price is justified, making cost estimates dependent on task scope, so without clear scope parameters, you'll face unpredictable bills. - Sonnet 5.5 runs $2 to $10, a sensible default for most tool-using steps. - Haiku 5.5 runs $0.10 to $0.50, so high-volume steps such as classification or routing cost pennies. Start on the cheapest tier that passes your test cases, because the model choice moves run cost far more than the platform plan does. Before you commit, confirm these at the [pricing page](https://www.activepieces.com/pricing): - MIT Community Edition and self-hosting terms - Free and Plus plans, including the flat $20 Plus price - SSO, audit logs and embedding, which go through sales - SOC 2 status ## Common problems with AI agents Start with the sceptic's case, because it is strong. **Agents vary between runs**, fail without raising errors, cost more per run, can be hijacked through their tools, and are often unnecessary. For non-determinism, fix the steps around the agent and validate its output against a schema. For silent failure, log every run and alert on empty or malformed results. For higher cost per run, cap iterations and tokens, and price a run before enabling it. For tool access and prompt injection, grant only the tools the task needs, read-only where possible. At scale, few organisations have tested this. In [Capgemini's survey](https://www.capgemini.com/insights/research-library/ai-agents/) of 1,500 organisations, **61% are still exploring agents** and 23% are running pilots, so most organizations won't have a local success story to emulate.  Only 12% have reached partial scale, and 2% operate at full scale, which means scaling agents is still an exceptional achievement. ### Inconsistent results and silent AI agent failures On different days, the same input can produce different tool calls. Worse, a wrong answer looks like a right one: no error, just a plausible record written to the wrong place.  Where a fixed step throws an exception, an agent returns confident text. So check outputs inside the workflow, route low-confidence cases to a human, and keep run history you can replay. If an agent reads untrusted messages and can also send email, it is a data-leak path, because every tool is something an attacker can steer with text in an email or support ticket. ## How to avoid agent mistakes Trace nearly any failure back and you find one move: someone took the fixed workflow away from the agent and let it improvise across the whole job. | Mistake | What goes wrong | Fix | |---|---|---| | Too many tools | The agent picks the wrong action, or an adjacent one | Give it only the tools this one step needs | | Vague instructions | Output shifts between runs and breaks downstream mappings | Name the task, the allowed outputs, and the stop condition | | No approval on irreversible actions | A wrong email, refund, or deletion goes out before anyone sees it | Put a human approval step before the action | | One happy-path test | Edge cases appear first in production | Test empty, malformed, and hostile inputs | Sending and paying can't be taken back. **Place an approval step ahead of each one, so the agent proposes and a person confirms.** Reversible actions, such as writing a draft or tagging a record, can run unattended. ### Judging the agent on one good run Feed the step an empty field, a malformed payload, and an input that tries to redirect it, then read the run logs for what it did with each. On Monday, pick one workflow, cut its tools to the minimum, and add one approval step. [Start free](https://cloud.activepieces.com/sign-up) to build it, or see [the pricing page](https://www.activepieces.com/pricing) for plan limits. ## Frequently asked questions ### Is Activepieces free to self-host for AI agents? Yes, the MIT-licensed Community Edition is free to self-host, and you can run agent steps inside your own flows on it. What it doesn't include are the paid-plan features, such as SSO beyond the Team plan and audit logs. If you need any of the paid features on your own infrastructure, that's a conversation with sales, not a setting you flip. Check the pricing page for what each plan includes before you commit to a rollout. ### Can an Activepieces agent use any app I connect? No. An agent can only call the tools you hand it, and each tool is a specific integration action, such as one trigger or one action with its fields mapped. A connected app isn't automatically an agent-usable app. Before you build around an app, read what actions its integration exposes and whether they cover your case. A long catalogue says little about whether the one action you need exists. Does Activepieces hold SOC 2 certification? Not yet. The SOC 2 programme is in progress, so don't write "SOC 2 certified" into a vendor questionnaire or a procurement form. If your security review requires the report today, ask sales where the programme stands, and consider self-hosting so your data stays in your own environment while you wait. ### Which language models can an Activepieces agent use? That depends on which AI providers you connect. The list changes as vendors release models, so check the agent step's model setting in your own workspace rather than trusting an article's list. Pick the model per step. A cheap, fast model suits a classification step, and a stronger reasoning model suits a step that has to plan across several tool calls. Because the agent is one bounded step, you can swap the model and see the effect in the run history. ### When should I talk to sales instead of starting free? Talk to sales when you need self-hosted enterprise, SSO beyond Team ($200 a month, 25 users included), or audit logs, or when you want to embed [Activepieces](https://www.activepieces.com/pricing) in your own product, which starts from $36,000 a year. Free, Plus and Team are self-serve, so only Embed, Ultimate (which adds audit logs and SSO beyond Team) and self-hosted enterprise control go through sales. If you are one builder or a small team wiring up your first agent step, [start free](https://cloud.activepieces.com/sign-up) and see whether the flow holds up. For plan limits, see [the pricing page](https://www.activepieces.com/pricing). For the enterprise items above, use [the sales form](https://www.activepieces.com/sales). ## Related reading - [Build Fillout Forms Automations with Activepieces (2026)](https://www.activepieces.com/blog/building-more-powerful-forms-fillout-form-activepieces) - [5 MCP Flows You Can Build With Airtable MCP on Activepieces (With Templates)](https://www.activepieces.com/blog/5-mcp-flows-you-can-build-with-airtable-mcp-on-activepieces-with-templates) - [SAP Business One AI Agents: Build a Chat Automation](https://www.activepieces.com/blog/sap-business-one-ai-agents-build-a-chat-automation) ## References - [Anthropic](https://platform.claude.com/docs/en/about-claude/pricing) - [Capgemini](https://www.capgemini.com/insights/research-library/ai-agents/)Every serious approach converges on the same rule: let the model propose and let a person or a deterministic check commit.