Agentic workflow platforms in 2026: the map, drawn honestly

Two reality checks before the map

In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Two months later, the same firm predicted 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025. Both predictions can hold at once. Agents are getting embedded everywhere, and a large fraction of the projects carrying them will die anyway.

The second reality check is more concrete. OpenAI launched AgentKit — a visual Agent Builder, ChatKit, Evals, and a Connector Registry — at DevDay on October 6, 2025. In June 2026, it posted a deprecation notice for the hosted Agent Builder, with shutdown on November 30, 2026. Roughly eight months from launch to sunset notice, from OpenAI, of all vendors. The migration path is self-hosting your own backend with the ChatKit Python SDK and the MIT-licensed Agents SDK; developers in the deprecation thread say that for small projects the migration economics do not work.

Read those two together and the lesson is not “avoid agents.” It is: the workflows you build will likely outlive the surface you build them on, so choose the way you would choose infrastructure — by license, by data model, by exit path — not by demo quality. That is the lens for the rest of this guide.

The five families

Most “top 10 agent platforms” lists compare products that are not substitutes for each other. The market sorts more usefully into five families, each optimizing for a different thing.

Family 1: automation platforms that added agents

These products existed before LLMs and earn their keep on integration breadth. The agent features are additions to a working automation engine, which is both the strength (mature connectors, mature scheduling) and the tell (agents inherit the platform’s short-lived, per-event execution model).

Zapier has the largest catalog in the category: the homepage claims 9,000+ connected apps, and Zapier MCP exposes 9,000+ apps and 30,000+ actions to MCP clients at 2 tasks per tool call. Its Human in the Loop app pauses a run for reviewers to approve, decline, or edit data, including guest reviewers via secure link. The honest limitation, per MIT’s AI Agent Index: Zapier Agents have no persistent memory between sessions, and no published safety evaluations or third-party audits were found. As of December 2025, the Agents product had only reached “the final steps of SOC 2 readiness” — the core platform is attested, the agent surface was still catching up.

Make connects 3,000+ apps with a paid plan starting at $9/month for 5,000 credits as of July 2026. Its rebuilt AI Agents reached general availability on February 11, 2026, with agents living directly in the scenario canvas and a real-time reasoning panel — a genuinely good building experience, and under six months post-GA as of this writing. The limitation that matters for governed work: the native Human in the loop app is Enterprise-only, so everyone else hand-rolls approvals through Slack buttons.

n8n is the engineer’s pick of this family: source-available under the Sustainable Use License (free self-hosting for internal business use; reselling it as a service requires a commercial license) and billed per full workflow execution rather than per step, which favors complex multi-step flows. SAP invested at a $5.2B valuation in May 2026 — more than double the October 2025 Series C. The cost of self-hosting showed itself between December 2025 and January 2026, when n8n disclosed four critical vulnerabilities in roughly two weeks, including a CVSS 10.0 unauthenticated remote takeover with public exploits. Self-hosting means owning that patch clock.

The rest of the family, briefly. Workato markets Enterprise MCP servers connecting agents to 14,000+ business apps (a vendor claim) but publishes no list prices — budget for a sales cycle. Tray.ai meters usage-based “Tasks” across 700+ connectors, quote-only. Pipedream bills one credit per 30 seconds of compute at 256MB and signed an agreement in November 2025 to be acquired by Workday — more on that below. Retool Workflows bills agents by wall-clock hour, with 20 free agent hours per month, but its workflows enforce a hard 15-minute run timeout that batch-job users hit repeatedly — a clean example of a family-level assumption (automations are short) surfacing as a product ceiling.

Shop here if your builders are ops people rather than engineers, your work is event-driven and short-lived, and connector breadth is the binding constraint. This family is genuinely good at that, and pretending otherwise would be lying to you.

Family 2: agent frameworks and their platform arms

Code-first frameworks for teams who want to own agent state, prompts, and control flow — each now paired with a commercial platform that monetizes deployment and observability. The recurring pattern to check: the framework’s license and the platform’s license are not the same thing.

LangGraph (Python and TypeScript) is the most established. The framework is MIT-licensed, durability is checkpointer-based — graph state persists per thread, enabling resume-after-failure and time travel — and first-class interrupt APIs support multi-day human approval flows. The catch sits one layer up: the production Agent Server, langgraph-api, is licensed Elastic-2.0, not MIT, and self-hosting it in production requires a commercial license key. Free self-hosting covers the framework; the managed-runtime experience is a paid product. LangSmith, the observability arm, runs $39/seat/month on Plus with traces billed from $2.50 per 1,000 as of July 2026.

CrewAI pairs an MIT-licensed framework (Crews for role-based agent teams, Flows for event-driven control) with the AMP platform. The vendor claims roughly 2 billion agentic executions in the trailing twelve months and customers including PwC and the US Department of Defense — vendor figures, not audited. Two honest frictions: the public pricing page shows only a free tier at 50 workflow executions per month and a custom Enterprise tier as of July 2026, with detailed paid tiers visible only behind a login; and production developers have reported the abstractions turning opaque — “you can’t clearly see what prompts are actually being passed to the LLM” — criticism that largely predates the built-in tracing added in the 1.0 release.

Mastra is the TypeScript entry, notable for moving in the developer-friendly direction on licensing: it relicensed from Elastic to Apache-2.0 in July 2025 after enterprise pushback, and launched its managed platform in April 2026. Two caveats worth weighing. Its durable-execution story still leans on third parties — the docs mark the Temporal integration experimental and point production users to Inngest. And on June 17, 2026, 145 npm packages in the @mastra namespace were compromised for 88 minutes via a stale ex-contributor account, in an attack Microsoft attributed to a North Korean group. Mastra responded quickly, but supply-chain exposure in this family is documented fact, not hypothesis.

The OpenAI Agents SDK rounds out the family: MIT-licensed with 27,646 GitHub stars as of July 2026. It survives its platform arm — the hosted Agent Builder is what is being shut down, and the SDK is the designated migration target.

Shop here if you have engineers who want control and are prepared to own serving, the approval UI, tenancy, and operations. The frameworks are free; the production posture is not.

Family 3: LLM-app builders

Visual builders whose center of gravity is the LLM application itself — chatbots, RAG pipelines, prompt-orchestrated apps — rather than back-office integration.

Dify is the heavyweight: ~148,000 GitHub stars as of July 2026, RAG as a first-class citizen (a visual Knowledge Pipeline and 30+ configurable vector store backends), and — since release 1.13.0 in February 2026 — a native Human Input node with durable pause/resume backed by Celery workers. Two things to read before committing. The license is a modified Apache-2.0: free single-tenant self-hosting, but you may not operate it as a service for multiple tenants without a commercial license, and you may not remove the frontend branding. And in June 2026, researchers disclosed four cross-tenant data-exposure flaws, two rated CVSS 9.1 and 9.4, patched in release 1.14.2.

Flowise (Apache-2.0 core, with the enterprise directory under a separate commercial license) was acquired by Workday in August 2025. Production self-hosters report progressive worker memory accumulation that is never released across job cycles. Langflow (MIT core, roughly 151,000 stars as of July 2026) now sits inside IBM’s watsonx portfolio via the DataStax acquisition announced February 2025. Its cautionary tale: CVE-2025-3248, a CVSS 9.8 unauthenticated remote code execution flaw that landed on CISA’s Known Exploited Vulnerabilities list and was used to build a DDoS botnet from exposed instances. Never run these public-facing and unpatched.

The hosted end of this family targets non-engineers. Gumloop (Pro at $37/month with 20k+ credits as of July 2026) has unusually strong compliance for its size — its trust page states SOC 2 Type II, HIPAA BAAs, and zero-data-retention agreements with major LLM providers — though reviewers report AI-heavy flows can exhaust a month’s credits in under a week. Lindy (no permanent free tier; Plus at $49.99/month) markets draft-and-approve human control on every tier, while an independent reviewer describes fast credit burn and “credit anxiety” that discouraged experimentation. Relevance AI ($24M Series B in May 2025) sells no-code multi-agent “workforces”; its official pricing page publishes no dollar figures, and the commonly quoted numbers — an Actions-plus-Vendor-Credits split with a Team tier around $234/month annual — come from third-party reporting, not the vendor.

Shop here if the deliverable is an LLM app — a support bot, an internal RAG portal — and speed to something working beats depth of workflow governance. Self-hosters get real data control and a real hardening burden; hosted-tool buyers should model credit consumption before committing a team to it.

Family 4: hyperscaler runtimes

The clouds sell agent infrastructure the way they sell everything: consumption-metered, deeply integrated with their identity and networking stack, indifferent to which framework you bring.

AWS Bedrock AgentCore reached general availability in October 2025 across nine regions. Its pricing is the purest in this guide: no platform fee, $0.0895 per vCPU-hour plus $0.00945 per GB-hour billed on active processing only, with per-component metering for gateway and memory — as of July 2026. The counterweight: researchers disclosed in March 2026 that the Code Interpreter sandbox permitted outbound DNS queries usable for command-and-control and data exfiltration, bypassing expected network isolation, and builders report operational rough edges such as memory contents being inspectable only through SDK calls, not the console.

Google consolidated Vertex AI Agent Builder and Agentspace into the Gemini Enterprise Agent Platform at Cloud Next 2026 — a consolidation-and-rename that is itself a small platform-risk data point. The substance underneath is strong: the open-source Agent Development Kit reached stable 1.0 across Python, Go, and Java, and the A2A interoperability protocol moved under Linux Foundation governance, with Google claiming 150 organizations running it in production. One planning note: exact Agent Engine compute rates were not verifiable on the official pricing page at research time — budget from a sales conversation, not the docs.

Microsoft Copilot Studio bills through prepaid $200 packs of 25,000 Copilot Credits or pay-as-you-go; its standout economics are that usage by Microsoft 365 Copilot-licensed employees is included at no extra charge. Two things its own documentation tells you: generative answers can include content from websites other than the one selected, with no accuracy check performed, and at 125% of prepaid capacity, custom agents are disabled until capacity is added — an overage policy that turns a budgeting miss into an outage.

Shop here if your organization is standardized on one cloud and procurement, identity, and network posture outweigh product polish. That is a legitimate trade, and for M365-saturated companies Copilot Studio’s included-usage math is hard to argue with.

Family 5: durable execution, and governed platforms built on it

The fifth family starts from a different question: not “how do I build an agent?” but “how do I make multi-step work survive crashes, deploys, and week-long waits?” That property — durable execution — is what most agentic failures in production trace back to lacking.

Temporal is the reference implementation. The server is MIT-licensed; workflows recover from crashes by replaying event history, retries and timers are durable, and a workflow waiting on human approval works identically “whether the wait is five seconds or five months” because timers persist in the server database. The company raised $300M at a $5B valuation in February 2026, explicitly positioned around agentic AI. Its honesty about scope is admirable: Temporal’s own agentic-control-plane post says enterprises still provide the agent logic, tool integrations, approval policies, and context preparation themselves. There is no builder UI, no connector catalog, no approval inbox. Users call workflow versioning its hardest developer-experience problem, and a third-party guide estimates realistic self-hosted production infrastructure at $2,500–$4,500 per month plus a fraction of an engineer. Temporal Cloud starts at the greater of $100/month or 5% of usage as of July 2026. If you have the platform team, this is the strongest foundation in the market.

Brahmalabs — and here we flag clearly that this is self-description; we build it, and you should read our head-to-heads where each claim sits next to the competitor’s counterclaim — is a governed platform built on that same substrate. Every run is a Temporal workflow, so a run that started yesterday can finish today; pause, resume, and cancel are signals, and approval waits park durably without burning compute. On top of the durability we build the parts Temporal says you build yourself — most visibly human approval as a first-class node (single approver or N-of-M consensus, SLA timeouts, escalation chains), per-run network egress control with the run page showing the domains each run actually reached, and an append-only audit log; the full feature-by-feature case lives in the head-to-heads. Pricing is credit-metered — Free at $0, Pro at $299/month, Enterprise custom — with bring-your-own model keys and no inference markup.

The candor owed in a field guide: Brahmalabs is a young product in early access, with no public case studies and SOC 2 Type II in progress, not certified. There is no self-hosted edition today — if self-hosting is a hard requirement, n8n, Dify, or Temporal’s open-source server are honestly better fits. Our integration catalog spans hundreds of providers, not Zapier’s thousands, though MCP support in both directions closes part of that gap. The head-to-heads against n8n, LangGraph, and Zapier go through these trade-offs product by product.

Shop here if your workflows are long-running, approval-gated, or audit-bound — the shape of work where per-event automation tools time out and frameworks leave you building the governance layer yourself. Temporal directly if you can staff the build; a governed platform on top if you cannot.

The consolidation signal

Three of the best-known open and independent agent-building tools now sit inside enterprise-software giants. Workday acquired Flowise in August 2025 and then signed to acquire Pipedream in November 2025, explicitly to give its agents connector reach. IBM’s acquisition of DataStax, announced February 2025, pulled Langflow into the watsonx portfolio. Add SAP’s investment in n8n at a $5.2B valuation and OpenAI’s Agent Builder shutdown, and a pattern emerges: the independent middle of this market is being absorbed or abandoned.

For a buyer this is not automatically bad — acquirers bring stability and budgets. But it changes the diligence question from “is this product good?” to “what happens to my workflows when this vendor’s owner, roadmap, or license changes?” Open cores with permissive licenses (Langflow’s MIT, Flowise’s Apache core, Temporal’s MIT server) give you a fork-and-run answer. Hosted proprietary surfaces give you a migration project on someone else’s timeline.

A seven-question selection rubric

Answer these in order. Most teams eliminate three families by question four.

1. Is self-hosting a hard mandate? If yes: n8n (free for internal use under its Sustainable Use License), Dify (single-tenant only, per its license), Flowise, Langflow, Temporal’s MIT server, or the CrewAI/Mastra frameworks. Note the fine print: LangGraph’s framework is MIT but its production server is Elastic-2.0. Zapier, Make, and Brahmalabs are managed-only.
2. Who builds — engineers or ops? Engineers: families 2 and 5. Ops and business teams: families 1, 3, and 4. Mixed teams: Dify and n8n sit closest to the boundary.
3. Is the work long-running or approval-gated? If runs span hours to weeks, or humans gate steps, you need durable execution (family 5), LangGraph’s checkpointer-plus-interrupts, or Dify’s post-1.13 pause/resume. Check ceilings elsewhere: Retool’s 15-minute timeout and Make’s Enterprise-gated approval app are family-level tells.
4. Do you have data-residency constraints? Self-hostable options solve it structurally. Among managed platforms: Make offers an EU data center, LangSmith runs EU and APAC regional instances, and Brahmalabs defaults to EU residency. Zapier hosts on AWS in the United States with no EU-residency option.
5. Do you need certification today? SOC 2 Type II holders you can verify include Gumloop, LangSmith, CrewAI, Temporal Cloud, Make, and Dify. In progress and honest about it: Brahmalabs, and — as of December 2025 — Zapier’s Agents product specifically. If the audit is this quarter, “in progress” is a no.
6. Can you predict the pricing meter from your workload? Per-execution (n8n) rewards complex flows; per-task (Zapier) punishes them; credits (Make, Copilot Studio, Gumloop, Lindy) are the meter users most often report being surprised by; pure consumption (AgentCore, Temporal Actions) is predictable if — and only if — you can model your compute. Run your three ugliest workflows through each meter before signing anything.
7. What is your platform-risk tolerance? After AgentKit, treat this as a real axis. Prefer permissive-licensed cores you could fork, exportable workflow definitions, or infrastructure with an ecosystem of operators. A proprietary hosted builder from a fast-moving vendor is a fine prototype surface and a risky system of record.

Where this leaves you

If you are an ops team wiring SaaS tools together and adding an AI step or two, family 1 is your home, and Zapier or Make will serve you well within their execution models — go in with eyes open about task and credit math. If your engineers are building agents into a product, start in family 2: Mastra if you live in TypeScript, LangGraph if you need battle-tested state control in Python, CrewAI if role-based multi-agent structure fits the problem — and budget for the serving and governance layers the frameworks leave to you. If the deliverable is an LLM app with retrieval, Dify is the most complete self-hostable option in family 3, provided you respect its license terms and patch cadence. If your company’s cloud strategy decides for you, family 4 works, and Copilot Studio’s included-usage economics are genuinely compelling for M365 shops.

And if your workflows are the long-running, human-gated, audit-bound kind — the kind Gartner’s canceled-project statistic is quietly about — family 5 is where the durability lives. Build on Temporal if you have the platform team to own workers, versioning, and the approval surface. If you want that substrate with the governance already assembled — approvals, sandboxing, egress control, audit — that is the bet we made with Brahmalabs, and the head-to-heads linked above are where we argue it against each alternative, gaps included.