Coordinating specialized agents creates a different mess. Rasa explains how orchestration is supposed to keep those agents from working at cross purposes. Deloitte’s 2026 technology predictions are more concerned with what happens when that fails: agent sprawl, spread across frameworks, languages, and infrastructure, with no clear owner.
Building these systems will get easier over the next few years. I do not think easier building is the interesting part. Visual tools keep attracting teams that do not want to manage infrastructure, and that is a sensible way to get a first workflow up. It tells you less once the workflow is simply how a task gets done each day. A tool call fails and, from the outside, you cannot tell what the agent already did. Keep it running as part of the working day and the cost stops being a surprise you can shrug off. Someone also asks where the data went. A framework with plain answers on those points is easier to live with. A tool that only has to look good in a demo, and then leaves the hard part with the customer, is a poorer fit.
The center of gravity is your documents. Agents and retrieval-augmented apps run over internal files in Python and TypeScript, and a Workflows module that reached 1.0 coordinates multi-step and multi-agent processes. LlamaAgents packages them as deployable document agents. This makes LlamaIndex a natural candidate when retrieval is central to the product, but it is not a substitute for a recovery strategy: unlike LangGraph, the setup described here does not include durable execution or a visual builder. The open-source framework for document-grounded agents is free under the MIT license. LlamaCloud adds a hosted layer with a free credit allowance, and paid plans start from per month.
Execution is modeled as an explicit graph with checkpointing, so a stateful agent can pause, resume, and recover after failure. Where CrewAI organizes work as role-based crews and omits durable execution, this stack treats recovery as part of the graph. There is no visual builder. The open-source framework for controllable multi-agent graphs is free to self-host with no license fee. LangGraph Platform adds a managed layer. A free Developer plan covers self-hosted deployments up to 100,000 nodes executed a month.
The practical unit is a crew: named roles and tools, with tasks handed between autonomous agents in Python. That is a different starting point from a node canvas or an explicit state graph. The open-source multi-agent orchestration engine is MIT-licensed and self-hostable with no license fee, so spend is compute and model tokens. AMP sits on top as a hosted no-code builder with monitoring and execution management, priced on usage above a free Basic plan.
Website: learn.microsoft.com
CrewAI
By Randy Ferguson
How many agents a framework can coordinate is a poor way to choose one. Start with the job: a visual builder for a quick workflow, a document-focused stack for internal knowledge, or code-first tooling when control and recovery justify the engineering work. Then test recovery, visibility, and whether the workflow can move if costs or requirements change. A person still has to explain what the system did.
Overview
A visual canvas is the default path, wrapping prompt orchestration, retrieval pipelines, agents, and model management in one product. Where CrewAI’s free hosted layer is measured in workflow executions, Dify’s Cloud Sandbox is free for 200 message credits. Paid Cloud starts from per month. The visual, self-hostable LLM application platform also ships an Apache 2.0 community edition that is free to self-host with no usage cap. That edition has drawn over 80,000 GitHub stars, a useful adoption signal but not a measure of production reliability.
Considerations
Credit-priced Cloud can diverge fast from the unlimited community edition, so volume should decide which edition you actually run. That is a different meter from CrewAI’s per-execution AMP free tier.
Checkpointing is built into the graph, so a long-running agent can stop and continue without replaying every node. Enterprise plans offer cloud, hybrid, or fully self-hosted deployment of the platform layer. Plus is per user per month and bundles LangSmith Plus, a figure current in October 2026.
Dify
Recovery after failure is still your problem, unlike LangGraph’s checkpoints, and there is no visual builder. Listed compliance includes SOC2, HIPAA, and GDPR, which matters if document agents touch clinical files. LlamaCloud bills in usage credits on top of the free MIT framework, so keep those lines separate when you model spend.
A demo does not have to survive a normal Tuesday. Customer requests do. They come in badly worded, they come in again the next morning, and someone is waiting on the answer. Internal documents are awkward for a different reason. Weeks later a colleague asks what the agent was allowed to open, and you need a real answer. A business process is stricter than either of those. The result does not stay in the chat. It goes to another person, who then wants to know what happened. Once a company hands that kind of work over, a failed run stops being a small annoyance. So does an action nobody asked for. A bill you cannot tie to a run is worse, because someone in finance sees it before the team that built the agent does.
Overview
Website: crewai.com · pricing
Key Capabilities
LLM steps are the native blocks on a visual node builder, aimed at research and enrichment pipelines. There is no open-source engine and no self-hosted edition, which is the break from CrewAI and Dify. The AI-native workflow automation platform is hosted only. A free tier includes 5,000 credits a month. Paid plans start at per month, a figure current on the company’s pricing page in October 2026.
Considerations
Best suited to: Growth and research operators automating AI-heavy pipelines without writing code or running servers.
Most startups do not need a sophisticated set of agents working together. One model is enough for a lot of this. A straightforward workflow covers more than a product page makes it seem. Ordinary software, with no agent in it, is still the better choice when the task barely changes and undoing a bad step costs real money. Every agent you add is another part that can break. If it is not worth the trouble, you have only given the team more to look after.
Gumloop
Demos skip the freeze. A startup puts an agent on prospect research, or on customer requests, and one bad tool call can stall the whole run. Starting again from the middle is not guaranteed. A log a person can read is not guaranteed either. Do this thousands of times and two dull questions get urgent: who is paying for it, and who is holding the data it stored on the way.
Startups change direction quickly. They switch models, replace a supplier, and rethink the product, sometimes inside a single quarter. The change itself is ordinary. A workflow that cannot move with it is not. Build the workflow around one platform and leaving often means rebuilding everything. Open-source frameworks are the direct way around that, if the team will run them on its own infrastructure. Hosted can still be the right call. The brochure will say convenience. Check whether the runs finish, and whether the hours you saved are still saved a few months later. If the pricing stops making sense, moving the work elsewhere cannot mean writing it all again.
Overview
We would be wary of treating the trial as a cost baseline. Seat charges begin when the week ends, with no self-hosted exit of the kind CrewAI and Dify still offer.
Considerations
Website: gumloop.com · pricing
Website: llamaindex.ai · pricing
LangGraph
For a startup, the easiest build tool is a weak reason to pick one. The workflow has to be something the team can maintain. It has to recover when it breaks. And if the product or the bill stops making sense, the team has to be able to walk away. The seven options below answer those questions differently. A visual builder suits teams prioritizing speed; a code-first framework offers more control over execution. Document-heavy projects have their own needs. These are starting points, not substitutes for checking recovery, data handling, and the cost of leaving.
Pricing and free-tier allowances below are snapshots recorded in September and October 2026, where dates are given. Vendors can change them. Credits, executions, nodes, and per-seat charges are different measures, so compare the workload each plan actually covers rather than the headline numbers alone.
Overview
There is no shared recipe. Some frameworks are open source and run on a company’s own infrastructure. Others are hosted, with a visual builder, so people can string a workflow together without writing much code. Convenience versus control is the short version. It is also a bit too neat. The products split in messier ways once you get past the marketing page.
Key Capabilities
Ask an agent a question and it will answer. That part is easy. Research is harder: it has to go look things up, pull the useful bits together, compare them, and write something you can actually use. A second agent might check that work, or just pick up the next task. None of this runs on its own. A framework has to sort the handoffs, decide which tools get called, and do something other than freeze when a step fails.
Considerations
Best suited to: Python teams that want role-based multi-agent crews they can self-host from day one.
Lindy
The SDK covers Python and .NET, succeeding Semantic Kernel and AutoGen after Microsoft moved those projects to maintenance. The 2026 release checkpoints workflow state so a run can resume later. Where LangGraph puts checkpointing in the graph itself, a durable extension here runs that recovery on Azure Functions or on self-hosted infrastructure. The open-source SDK for multi-agent workflows is free. Azure-hosted agents are billed by consumption, a usage model last verified in October 2026.
Website: langchain.com · pricing
Overview
Considerations
There is no visual builder, so non-engineers are unlikely to own the graph directly. LangGraph is worth examining when a multi-step process must resume from saved state after a failure and the team is prepared to write and maintain the control flow in code.
LlamaIndex
Best suited to: Engineering orgs that need checkpointed, recoverable multi-agent graphs and will write the control flow in code.
Website: dify.ai · pricing
Overview
Considerations
Hosted visual agents handle scheduling and email triage for people who will not write Python. Unlike Gumloop, there is no standing free tier, only a 7-day trial. Plus then starts at .99 per user per month, a figure current in October 2026. The no-code AI agent builder has no open-source or self-hosted edition, and nothing to install on your own machines.
Best suited to: .NET and Python groups that want a free SDK with checkpointed agents on Azure Functions or their own hosts.
Microsoft Agent Framework
Retrieval pipelines live on the same canvas as agent steps, so a document-grounded workflow does not require a second product. Model management sits there too, which keeps prompt edits and model swaps in one place. Listed compliance is SOC2, ISO27001, and GDPR. Cloud Team is 9 a month. That plan includes 10,000 credits, with custom Enterprise pricing above it.
Overview
Best suited to: Non-technical founders deploying email and scheduling agents through a hosted visual builder.
Key Capabilities
You cannot self-host, so prompts and runs stay on the vendor’s platform. That may be a reasonable trade for a small team that values speed over infrastructure control, but it rules Gumloop out where deployment location is a hard requirement. Dify’s community edition and CrewAI’s MIT engine offer different routes for teams that need to run more of the stack themselves.
Considerations
It works with models from multiple providers, which lets Python and .NET agents call more than one model vendor. Azure AI Foundry is the integration for hosted agents.
Best suited to: Product groups that prefer a visual LLM app builder with an unlimited self-hosted community edition.
When the Demo Becomes Real Work
There is no visual builder, so people who will not write Python or .NET should look at Dify’s community edition. LangGraph Platform meters a free Developer plan by nodes executed; Microsoft’s hosted agents have no such allowance and bill by Azure consumption.
Website: lindy.ai · pricing
Best suited to: Builders wiring agents to internal documents in Python or TypeScript under an MIT license.
The basic crew pattern does not provide the same built-in checkpointing as LangGraph, so teams should establish how retries and partial failures will be handled. AMP’s free Basic plan includes 50 workflow executions a month, a figure recorded in September 2026; that allowance may be thin if the team relies on the hosted builder. CrewAI Factory offers a containerized self-hosted path when AMP is no longer enough.
