McKinsey’s latest State of AI report revealed that 62% of organizations have started experimenting with AI agents in some form. However, Gartner expects over 40% of these projects to be scrapped by 2027. This is mainly because among the thousands of vendors promising to provide ‘agentic AI’. Gartner reckons that only 130 of them provide genuine agentic functionalities, while the rest are “agent washing.Both figures, despite seeming at odds, are true. Agents’ actual adoption is growing, as operational teams begin to use them to triage emails, enrich leads, or route tickets and that too without writing any code. At the same time, the inflated expectations around agentic AI are quickly meeting reality, and the disconnect is most evident in a place every business cares about – the price.AI agents are solving some of the most complex schedules with just a click. Rather than hiring a human who performs a repeatable procedure, one can simply connect the agent, who performs it instantly upon firing the trigger. It enriches the lead before the salesperson even gets to work, sorts tickets according to priority in the queue by 7 am, without expanding the team.Unfortunately, the term “AI agent builder” applies to such a wide spectrum of real reliability. While some platforms can handle a truly messy and complex workflow, others break after just three triggers or work perfectly until they get stuck in a loop and quietly charge you a fee that you didn’t approve of. However, pricing makes it quite complicated to evaluate a product before using it. Credits, activities, executions, conversations – whatever it is called on a platform – can mean something else on another. The cost displayed on the platform’s home page usually gives you little idea about what a monthly fee will be in real life. That’s why we unraveled the best eight platforms for you.Quick ComparisonPlatformBuild styleMetering modelFree tier (the honest version)BYOKSelf-hostableGumloopVisual canvasCreditsNone currently — 14-day free trial onlyYes (50% off credits)Non8nLow-code / canvasExecutionsSelf-hosted community editionYesYesZapier AgentsConversational + no-codeActivities400 activities/moNoNoLindyConversationalCredits (Undisclosed counts)7-day trial onlyNoNoDustPrompt-basedCredits per seat500 credits, one-timeN/A (Model choice included)NoRelevance AICanvas (“Workforces”)Actions + Vendor credits200 Actions/moYesNoMindStudioVisual canvasNone — At-cost tokens1,000 runs/mo, permanentYesBusiness tier onlyBotpressVisual + code (ADK)Conversations (AI bundled in)100 conversations/moEnterprise-gatedNo (legacy v12 only) How to Choose an AI Agent BuilderHere is the part that is rarely discussed in such comparisons, but is of utmost importance to anyone who wants to build agents that perform some repetitive, operational tasks.The metering decoder: Credits vs. activities vs. executions vs. conversationsCredits (Gumloop, Lindy, Dust): Most platforms that allow you to build agents use credits as the unit of measurement. Each action the agent performs costs a certain number of credits depending on the complexity. It can be a simple API call for an LLM prompt, a “deep” request using the larger models, web scraping, and data enrichment steps. The more complex the task the agent has to perform, the higher the number of credits for that step.Activities (Zapier): Similar to credits, but specific to agents; each step taken by the agent (trigger, knowledge base question, web search, or tool) adds up to the activities counter.Executions (n8n): As a counterbalance to execution-based billing (more on that below), n8n’s execution is a completed workflow run. Even if your agent has to run 10,000 operations in one go on another platform, in n8n, this will count as 1 execution.Conversations (Botpress): Botpress bills based on conversations, or any exchange of messages between the user and the agent that has at least 2 messages in it.At-cost tokens (MindStudio): An interesting tokenization-based option, with no added fees, that bills based on the exact number of tokens used with third-party models.None of these options is better or worse than the others, but bear in mind that the faster you plan to execute your agent, the more of any of these units it will bill for.Agents vs. classic automation platformsThe difference between agents and Zapier, n8n, or Microsoft Power Automate lies in the fact that in the latter, any automation is strictly deterministic. When an email comes in, it triggers a certain set of actions that are hard-coded when setting up the automation. An agent, however, has an LLM in the mix that allows it to make decisions during its execution based on what it sees – what lead to prioritize, what email to reply to, and whether an incoming ticket requires human input.Zapier itself suggests that the most efficient way to use its agents is to combine them with “deterministic” actions, reserving the use of LLMs for the points where human judgment is required. The debate between n8n’s executions and Zapier’s activities is therefore not unlike agents’ use cases and pricing – n8n’s deterministic, no-code automations allow businesses to build fully fledged, repeatable processes, whereas Zapier’s more complex billing better reflects the use of agents.No-code agents, low-code agents, and agents for developersEach company offering agent-building tools will claim, quite justly, that their product is intuitive and does not require coding to use. However, only two of the eight platforms we reviewed (Zapier Agents and MindStudio) are genuinely no-code in their current state and ready to be used by an average operations person on day one. n8n positions itself as a “low-code” automation platform – which it is – and requires some technical proficiency to string together individual API calls (or run them on a self-hosted instance behind the company’s firewall). Security and complianceIf your agent has access to your inbox, CRM, or other operational systems, it is, by virtue of that access, a security risk. Any agent platform worth its salt will have SOC 2 Type II compliance, and it is always a good idea to ask whether you can use your own encryption keys (BYOK) when setting up agents.Less obvious but equally important – check if there is a feature or a policy that would require human approval for any agent-initiated action. Such a safeguard is particularly important for security reasons, as demonstrated by the Repello AI team last year – by exploiting the lack of approval guardrails in Zapier-connected agents, the researchers were able to exfiltrate email content, including passwords, from user accounts over unencrypted channels.The 8 Best AI Agent Builders in 20261. Gumloop – The visual agent canvasYC-backed visual workflow builder, sometimes compared to “Zapier on steroids”, which focuses on two innovations: subflows (entire workflows as a single node) and subagents (agent spawning for parallel work). It bagged a $17M Series A in January 2025, followed by a $50M Series B round led by Benchmark in March 2026.Gumloop – the visual agent canvasIts customers include Shopify, Ramp, Instacart, and Gusto, according to its website.A typical call to an AI assistant is 2 credits, a call to a heavier model (GPT-4.1, Claude Sonnet) is 20 credits, and a step of enrichment is 60 credits per contact. Gumloop’s own documentation has an example of a workflow that enriches 100 contacts costing 6,000 credits, or more than a free tier’s lifetime budget in a single workflow.As of our last check, Gumloop’s pricing page doesn’t have a permanent free tier anymore. The 14-day free trial is the closest alternative. Pro starts from $37/month for 20,000+ credits with unlimited seats, and scales via a slider up to roughly $1,840/month for 1,000,000 credits; Enterprise is built on demand with SCIM/SAML, VPC deployment, and audit logs. If you’ve seen an article mentioning a 5,000-credit free tier, it may already be out-of-date. Gumloop is one of the most dynamic in its pricing. BYOK users get a 50% discount on AI model credits for Pro and higher.The Gumloop community forum has a couple of threads mentioning a production workflow mistakenly run on a 759-row CSV instead of a 5-row one, costing around 17,000 credits in a matter of seconds, and another user citing a 1,200-credit scrape with no results. On the security front, it is SOC 2 Type II, SAML/SCIM, VPC deployment, no data retention with model providers, and no HIPAA compliance.2. n8n – The self-hostable powerhouse (low-code)The Berlin-based workflow automation platform has an interesting licensing model: “fair-code”, or Sustainable Use License. This doesn’t actually make it open source, and forbids reselling it as a hosted service (while the code is available on GitHub and has nearly 199,000 stars). Its AI Agent node is built on top of LangChain, and its live integration directory lists over 1,900 apps.n8n – the self-hostable powerhousen8n raised a $180M Series C at a $2.5B valuation as of October 2025, with Accel as one of the participants.n8n’s pricing is the most straightforward among similar tools: a single execution of an entire workflow (regardless of its internal branching and data volume) is counted as 1 execution. It is actually explicitly compared to other tools that use per-step billing in its own documentation. The same fairness extends to its AI credits: 1 Agent call = 1 credit spent.Starter is €20/month (paid yearly) for 2,500 executions and 2,300 AI Assistant credits; Pro is €50/month for 10,000 executions and 13,700 credits; Business is €667/month for 40,000 executions with SSO, SAML, and version control; and Enterprise is custom. For the self-hosted Business plan, overage pricing is €4,000 for an additional 300,000 executions. Community Edition, the self-hosted open-source version, is free.OpenAI, Anthropic, Gemini, Bedrock, Groq, Mistral, local models with Ollama – you name it. You pay model providers directly for the usage, rather than n8n.n8n positions itself as a low-code platform, which it indeed is – but the AI Assistant is only one node in a complex visual builder. The actual coding is required to design a moderately sophisticated parallel workflow, and if you go self-hosted, you’ll have to maintain a Docker container. This is not a tool for a non-technical person, but rather an engineer who wants full-stack control.3. Zapier Agents – The easiest on-rampAn agent layer built over Zapier’s ecosystem of over 9,000 connected apps, designed to be built with natural language and visual help from Zapier’s Copilot. It is an independent company, and while it has a stellar product, it’s not exactly “big tech”. It’s been profitable since 2014 and has raised only 1.2 million in its entire history.Zapier Agents – the easiest on-rampActivities, which are similar to actions in other platforms, but with the critical difference that they’re not tied to a Zapier Zap. Every agent action (trigger, knowledge base query, tool call, web search) is counted as 1 activity. The free tier includes 400 activities/month; paid plans are said to start at around $33/month for 1,500 activities. However, Zapier’s website has a JavaScript-obfuscated agent pricing page, which makes it difficult to provide an exact quote.First, Zapier’s own knowledge base has an article about activities being unexpectedly consumed by an agent and a Zap triggering each other in a loop. Second, the July 2025 Repello AI security demonstration used a Zapier Gmail agent as its proof-of-concept automation, and the researchers note that a human approval step should be added before any agent touches sensitive data.It is the largest catalog of third-party tools by far, and it’s extremely easy to get started with a Zapier Agent if you’re already familiar with the platform. The free tier is generous enough to test an agent-assisted workflow.Activity-based billing on top of task-based Zap pricing can create budget surprises if you use the two together; agent pricing transparency is lower than with other tools. An established Zapier user looking to add agent-assisted automation to their existing Zaps.4. Lindy – The AI-employee approachA conversation-layer platform, which allows you to delegate tasks to individual “Lindies” – agents with fixed roles (replies to emails, meetings, etc), which can message each other to perform more complex actions.It has credits, but with a twist. Lindy doesn’t publish exact credit values for each plan. Instead, their pricing page only shows relative values (higher plan = more usage). It’s hard to overestimate how important this is: Gumloop and Dust, two similar tools, publish their credit values, and users can calculate the costs based on it.Lindy – the AI-employee approachLindy’s approach is arguably less transparent. Simple actions take between 1-3 credits on smaller models and around 10 on larger ones. Credits are not banked between agents, and agents are suspended when a plan is exhausted. One independent review estimates that a lead generation workflow would take 3 credits to search a knowledge base, 7 credits to send an email, and 265 credits to make a phone call, or 275 credits per lead.Lindy no longer offers a free tier; new users can try the product with a 7-day trial. The paid plans start at Plus ($49.99/month), move to Pro (reported between roughly $60 and $100/month depending on the source, reflecting a recent relaunch of the pricing page), and top out at Max ($199.99/month) before Enterprise, which is custom-priced and has an additional onboarding fee. Model access is gated behind a paid plan, and BYOK is not supported.The most compliance certifications of any agent platform. It maintains SOC 2 Type II, GDPR, HIPAA, and PIPEDA.Lindy’s most common complaints are about the value of credits. An agent is more expensive than a traditional Zapier Zap, and the product is not as broadly applicable as some competitors.It might be best for an independent worker or a small team looking to delegate narrow tasks (email management, meetings) to individual agents.5. Dust – Internal agents on company dataA Paris-based agent platform that allows building internal agents (for customer support, product development, etc.) that can search and reference internal data (Notion, Slack, Drive, GitHub) with citations. Its customers include Qonto, Alan, Datadog, and 1Password.It bagged a $16M Series A, which was followed by a $40M Series B in May 2026, led by Sequoia and Abstract with participation from Datadog, Snowflake, and others (according to Sifted). Dust’s recurring revenue is estimated at $20M across 3,000+ companies.Credits per seat is one of the four tools in this list that changed its pricing recently, and articles referring to a flat per-seat euro value are now outdated. As of our last check, a Free tier provides a one-time 500-credit allocation per seat, Pro is $30/month ($24/month paid yearly) with 8,000 credits per month, and a new Max tier is $150/month ($120/yearly for 40,000 credits per month. Enterprise is custom-priced, and programmatic/high-volume usage is billed at $0.01 per credit separately.It offers more than 20 models across all tiers, with no model locked behind a specific plan. One can pick a model for an agent, rather than a plan picking a model for you.It maintains SOC 2 Type II, GDPR compliance, and HIPAA-compliant data processing, with no data retention by model providers and regional data residency (EU/US).6. Relevance AI – the multi-agent workforce for GTM teamsAn SF/Sydney-based company that allows you to build GTM-focused “Workforces”: sets of agents that can be connected via AI or fixed logic to perform a larger task (e.g., sales outreach). Its GTM focus is similar to n8n’s process mining and Zapier’s Zaps, but it’s aimed specifically at GTM teams and has a unique approach to billing. It bagged a $24M Series B led by Bessemer in May 2025.Relevance AI – the multi-agent workforce for GTM teamsIt provides two separate meters, which is rare for an agent platform. Actions are counted every time an agent acts (tool call, web search, etc) inside a workflow, and separate Vendor Credits are spent on the underlying model, at a rate described as “no markup”. Unlike other platforms, which only count credits spent on models, Relevance AI’s Actions meter appears to bill for the execution of every step in a workflow, regardless of whether it’s a model call or a database query. However, unlike most other tools, Relevance AI’s Vendor Credits are persistent: you don’t lose them at the end of a billing cycle.Several competitive analyses suggested that Relevance AI’s self-serve tiers were being sunset in favor of enterprise sales only. As of our last check, this does not appear to be the case: the self-serve pricing is available, and has a Free tier (200 Actions/month plus a one-time 10 Vendor Credits), a Pro tier (roughly $19–29/month for 2,500 Actions and 10,000 Vendor Credits) to Team (roughly $234–349/month for 7,000 Actions and $70/month in Vendor Credits), and a custom Enterprise tier. BYOK is supported, which allows you to avoid Relevance AI’s Vendor Credits entirely.Reported overage pricing for Actions is around $40–80 per additional 1,000, which is higher than task-based alternatives, and the two-tier pricing model is more difficult to budget than a single Actions credit. The self-serve tiers have a limited feature set, and the review base is smaller than some competitors.7. MindStudio – The zero-markup budget pickA no-code visual agent builder that bills only for the platform usage and no model markups, no credits. Whatever you pay for a model, you pay for it in MindStudio, with no additional costs.MindStudio has no markup for models. It has a flat platform fee and passes along model costs for the Service Router (which has access to 200+ models without additional API keys) or your own keys for OpenAI, Anthropic, and Google.MindStudio – the zero-markup budget pickA permanent free tier (not a limited-time offer) includes 1 agent and 1,000 runs per month. Individual is $20/month ($16/month, paid yearly) and has no limits, while Business is custom-priced and includes team workspaces, SSO, audit logs, and self-hosting.It might be a steeper learning curve for complex multi-model workflows, with potential for costs to spike if a workflow isn’t optimized (no markup = no protection against higher model rates), and it is one of the few tools with limited third-party verification of its figures.It would be suitable for a no-code agency or a budget-conscious company looking to minimize model costs and have full control over their agent spending.8. Botpress – Customer-facing agentsA Montréal-based company that has transitioned from a general-purpose chatbot platform to a customer support and helpdesk solution with an emphasis on AI. It has a visual Agent Studio, a code layer called Autonomous Engine, and a human handoff.Its billing is done per conversation, rather than per message. It is the fourth tool in this list to have changed its pricing recently. As of May 2026, Botpress transitioned to conversation-based pricing, which appears to be lower than per-message rates. A conversation is any exchange of messages between a bot, a human, or a combination of the two.Botpress – customer-facing agentsIts pricing suggests a free (Pay-as-you-go) plan has roughly 100 conversations/month for new workspaces. Plus runs about $150/month, and Team is about $750/month. Enterprise is custom-priced and supposedly starts at several thousand per month for multi-year commitments. There’s one important detail: auto-renewal for overage conversations cannot be disabled.It maintains SOC 2 Type II, HIPAA, and GDPR compliance with third-party penetration testing.The open-source version of Botpress is real, but it’s the v12 codebase. The self-hosted version described in this article is not available. If you want to self-host, this is not the company to ask.Flow complexity can balloon expenses at scale; multiple conversations can cause increased latency if they have to make sequential LLM calls; bring-your-own-LLM is available but only at the Enterprise level.Relay.App Is Shutting Down – What That Says About This MarketRelay.app, a competitive alternative to Zapier, which was founded by a Google product lead behind Gmail and Calendar, announced its impending shutdown with no explanation on Tuesday, July 16, 2026. Free accounts will be closed on August 15, 2026, and paid subscriptions will be terminated on September 14, 2026. New user account creation and free-to-paid upgrades were disabled immediately upon the announcement. Overall, the company has raised $8.1M in two financing rounds.Firstly, note that many “best agent builder” lists from the first half of 2026 still feature Relay as an option. Always check the dates on articles or lists that compare agent-building platforms — the technology landscape has changed dramatically in the last 18 months. Secondly, the phenomenon of platforms dying in this space is an extremely tangible risk, and one should approach selection with eyes wide open. When choosing a platform for serious production use, consider self-hosting options (n8n), easy auditing of data flows (n8n), and an established funding runway in addition to your preferred UI and available features, as the combination of features alone has proven an insufficient foundation for long-term viability in this space.What Still Fails in ProductionWhile the demos often look impressive, it’s important to keep in mind that production capabilities lag far behind the marketing claims. Per the results of CMU’s “TheAgentCompany” benchmark (available here), which evaluates models on their ability to perform realistic tasks in an office productivity context, the best-performing model completed only 30.3% of tasks (Gemini 2.5 Pro; GPT-4o managed 8.6%). The examples of the tasks that failed ranged from fairly simple (“rename the user account to indicate completion”) to moderately complex.Gartner’s report on agentic AI corroborates this assessment, noting that over 40% of agentic AI initiatives are projected to be cancelled by the end of 2027, while the analysts also discuss the phenomenon of “agent washing”, i.e., vendors stretching the definition of an “agent” to include nearly any product. Their report notes that out of the thousands of vendors touting “agentic” offerings, only around 130 have products that demonstrably meet the definition.Meanwhile, the nuance provided by the MIT NANDA research is also important to keep in mind – while the researchers found that only around 13% of generative AI pilots in enterprises showed any positive financial impact, those utilizing purchased vendor tools had a 67% success rate as compared to 33% for internally developed solutions, a finding that could be applied to the tools reviewed here as well (and is another reason why the self-hosted n8n is a reasonable option for some).The credit blowup pattern is also something to keep in mind, as several platforms (Gumloop, Zapier, Botpress – each in different ways) feature a combination of a loop + credits/activities + automatic recharge, with Gumloop’s forums featuring a particularly detailed thread on a 5-row test that accidentally used 759 rows and burned 17,000 credits in the process of doing so. Zapier’s documentation even has an entry on agent-related activity usage blowups. Botpress documentation indicates that auto-recharge cannot be disabled, and Lindy’s documentation does not mention credits at all, which is part of the pattern that should be examined when trying to assess the risks. This serves as the most important lesson for any platform with visual no-code agent building: try it on a small scale and establish costs for a larger run before pressing the “go” button.Lastly, the security implications should also be reviewed — the Repello AI post on a Zapier-connected Gmail agent is yet another reminder that any agent that has access to email or CRM should be reviewed by a human being at all times. An agent that can read and write mail on your behalf is, at the very moment it exists, a security vulnerability regardless of the platform it is hosted on.The Big-Tech BuildersIf your team already works extensively within a single stack, there is a good chance that its major players already provide agent-building capabilities as either an add-on or part of their broader offerings. Google’s Gemini Enterprise, OpenAI’s new AgentKit (announced at their October 2025 devday), and Microsoft’s Copilot Studio each provide ways to build agents utilizing large language models, typically with significant pricing (per seat) and as part of a larger Workspace or Copilot offering. These are less flexible options when compared to the tools mentioned above, but if your team already utilizes any of Big Tech’s enterprise offerings, it is worth examining what additional features they provide before choosing an independent option.Should You Hire an AI Agent Development Company Instead?For many narrowly-scoped internal processes that require little to no customization, a no-code or low-code agent builder from the list above is likely to be the fastest and most cost-effective option available. The equation changes when complex custom logic, extensive internal software integration, or regulatory requirements are involved, in which case an agency or a dedicated in-house team working with raw models and APIs may provide better value in the long run due to the ability to tailor the solution to your needs specifically.If you are unsure which category your use case falls under, one simple test is to attempt to describe your desired outcome in a single paragraph without resorting to “… and if that fails, then…” statements – if you find that you cannot do so, it may be a sign that your solution requires a more tailored approach.Final Verdict: Match the Platform to Your Team’s CeilingThere are no right or wrong answers when it comes to choosing between the options, but it is important to understand what each can offer. If your team is uninitiated to no-code platforms, it makes sense to begin your exploration with Zapier Agents or MindStudio.If your team has engineers who want to build the most complex possible agent, Gumloop, Lindy, or Relevance AI are better options. If you need a full self-hosted stack with guaranteed SLAs to support business-critical operations, n8n is the only viable option. If your use case entails answering questions based on your company’s knowledge base, Dust is the best option available.Lastly, if your use case is customer-facing (support, sales, tickets), Botpress is the right choice. In all cases, run a small-scale test before going into production — the demos always look impressive, but it’s the 759-row version that will bill you for 17,000 credits.