TechnologyAgents & Tool Use
AI Agent Framework
At a glance
A library that provides the agent loop, tool calling, state and memory, multi-agent patterns and tracing hooks.
- Who this is for
- Engineers and technical readers learning the terms used in AI systems.
- Topics
- Agents & Tool Use
- Technology
In short
An AI agent framework is a library that provides the agent loop, tool calling, state and memory, multi-agent patterns and tracing hooks. Bigcircle uses one to assemble a prototype, then keeps only the parts that still earn their place: a bound on delegation, a budget on the whole run, a resume after a failed step, and a person on anything that cannot be undone. When role-and-task modelling no longer fits, we say so and port it off.
Key facts
Definition
An AI agent framework is a library that provides the agent loop, tool calling, state and memory, multi-agent patterns and tracing hooks, so you do not have to write that machinery from a chat completions call.
How Bigcircle does it
- Bigcircle has not ranked AI agent frameworks on a private benchmark, and none of its published case studies names the library, so the production decision is which parts to keep and which to replace with a provider SDK call.
- On a CrewAI crew, Bigcircle will not leave the run unattended until delegation has a ceiling, the budget covers the whole run rather than each agent, a stalled or looping crew is stopped, and a failed step resumes from a checkpoint.
- Bigcircle says when the job has outgrown the framework and ports it off, because role-and-task modelling stops fitting once the work needs durable execution, fine-grained human approval, or branching that is not a sequence.
Comparison
An AI agent framework ships the loop, tools, state and tracing, a harness is the runtime around one model that your team can debug, and orchestration is splitting work across specialists under a supervisor.
Key number
At 95 percent reliability per step, a 20 step run finishes clean about 36 percent of the time, which is why a framework that hides the step count still has to be bounded.
An AI agent framework is a library that gives you the AI agent loop, tool calling, a place to keep state, patterns for more than one agent, and hooks so you can trace a run. We reach for one when assembling a prototype from the provider SDK would take too long, and we keep only the parts that still earn their place.
The rest of the production work sits in code the model cannot edit: a bound on delegation, a budget on the whole run, a stop when the crew stalls or loops, and a checkpoint so a failed step resumes. The steps we actually ship live on our guide to building AI agents. Splitting work across specialists is agent orchestration, not a library choice.
We have not ranked the libraries on a private benchmark, and none of our published case studies names the one we used. The production decision is which parts to keep and which to replace with a direct SDK call.
What is an AI agent framework?#
An AI agent framework is the package that writes the boring machinery around a model: it asks the model for the next action, runs the tool, writes the result back, and stops when a condition says so. It usually ships a way to hold memory across steps, a pattern for handing work to another agent, and a trace you can open later.
The framework is not the agent. The agent is the model in that loop, with tools and a stop, and the framework is the code that makes the loop exist so you do not write it from the chat completions call. Microsoft's overview, checked 12 November 2026, puts the same fork in one line: if you can write a function to handle the task, do that instead of using an AI agent.
LangChain's docs, checked the same day, split the stack into three layers. A framework gives you the loop and the integrations. A runtime, which is where they put LangGraph, gives you durable execution, streaming, a pause for a person, and persistence.
A harness is opinionated: planning, files, subagents, prompts already chosen. We already sell that last layer as harness engineering, because a library's inner harness covers the common path and the work is usually in what it does not cover. Use the library when it shortens the path to a working loop, and drop to the provider SDK, or to plain code, when the abstraction is only costing you a stack trace.
That is the call our LangChain engineers page already makes: keep composition, streaming and retrievers, and call the model directly where the wrapper hides the failure.
How an AI agent framework actually helps#
Walk one request through the library. The model sees a goal and a list of tools, writes a tool call, and your process runs that call or refuses it. The result goes back, and the model picks again until the stop fires. Without the framework you still write that loop; the library saves you the first version of it, plus the callbacks that turn each hop into a span.
State decides whether the loop can leave the notebook. LangGraph's persistence page, checked 12 November 2026, writes a checkpoint of the graph so a crash, a timeout, or a wait for a person can resume. In-memory savers vanish on restart, so a store that survives the process is a production choice, not a default.
OpenAI's Agents SDK ships sessions for the same job, and a built-in trace so you can see the hops. Pydantic AI treats the model as a string you can swap, which makes the swap itself take a minute. Our model swap guide covers why proving the product still behaves takes a shadow run, a canary and evals.
A person still has to sit on anything that cannot be undone. An interrupt in LangGraph pauses the graph, shows the state, and continues after the edit. That is the same gate our human-in-the-loop guide already argues for, and it has to live in code the model cannot skip.
On Skionis an engineer approves a diagram before a cloud change applies. The library can pause, but it does not choose what the reviewer should see.
Our AI agent page already runs the arithmetic. At 95 percent reliability per step, a 20 step run finishes clean about 36 percent of the time. A framework that hides the step count does not change the product, so cap the run, trace every hop, and score the hops on your own cases. Our guide to building agents puts that eval work before the model choice.
What are the best AI agent frameworks?#
There is no best AI agent framework in the sense of a table you can paste into a purchase order. The useful question is which job the library still does after the prototype, and what you will have to put around it.
We pick on six things we already ship: durable resume, a place for a person, branching that is not a sequence, a trace you can eval against, how hard it is to call the provider SDK instead, and whether the model is a string you can swap.
LangChain's own list, checked 12 November 2026, names LangChain, CrewAI, the OpenAI Agents SDK, Google's Agent Development Kit and LlamaIndex as frameworks, and puts LangGraph with Temporal as a runtime. Read that as a map of layers, not as a ranking. We still name LangGraph here because people looking for an AI agent framework often need the checkpoint and the interrupt, which is the runtime column.
| Library | What its docs, checked 12 November 2026, say it ships |
|---|---|
| LangChain | The agent loop, model and tool abstractions, middleware |
| LangGraph | Durable execution, checkpoints, streaming, a pause for a person |
| CrewAI | Role, task and crew, plus Flows for state and control |
| LlamaIndex | Retrieval, indexes, and agents over your documents |
| OpenAI Agents SDK | A small set of primitives: loop, tools, handoffs, tracing, sessions |
| Microsoft Agent Framework | Agents, graph-shaped workflows, sessions, successor to AutoGen |
| Google ADK | Sequential, parallel and loop workflows, now pointing at graphs |
| Pydantic AI | A typed agent loop, with the model as a string you can swap |
CrewAI makes a multi-agent prototype quick to assemble, and that sentence is already on our CrewAI engineers page. What it does not give you is a bound on delegation, a way to stop a crew that has started looping, or state to resume from when one step fails. We put those in before anything runs unattended. A budget on each agent still lets a five-agent crew spend 5x what anyone approved, so the budget covers the whole run.
LlamaIndex is a retrieval toolkit that also ships agents. Most failures we see on that path start at ingestion, not at the agent loop: a parser that flattened a table, a chunk that lost its heading, metadata nobody extracted. Our LlamaIndex engineers page is the longer version of that. Do not pick it as "the agent framework" if the job is a loop over tools with no corpus.
What are agentic AI frameworks?#
People type "agentic AI frameworks" for the same libraries. Some of them own the path in code, and some of them let the model pick the next hop. We keep those as workflow and agent, and the agentic workflows guide is the page that makes that choice. An agentic AI framework is not a third category.
Google's ADK still documents sequential, parallel and loop workflow agents that run without asking a model what to do next. LangGraph lets you mix those deterministic nodes with model-driven ones. CrewAI's Flows hold state and control, and its Crews are the role-playing team inside a step. OpenAI's Agents SDK runs a loop until the task is complete, and a handoff moves the conversation to another agent. Those are different designs under one search.
If a person can write the hops, write them. Graph engineering is that advice in another vocabulary, and multi-agent orchestration is the page for a lead that splits work. Do not buy a multi-agent crew because the README showed four named roles. Most work that looks like five agents needs two and a better tool, which is how our orchestration service starts.
On Solarpunk the next tool depends on the last result, so we kept one loop and a critic that decides continue, retry or escalate. Plans completed with no person rose from about 55 percent to 80 percent once local retrieval and planner memory grounded the agent in the user's workspace.
A framework that wants a fixed crew would have split tools the agent has to discover as it goes. That is the case that needed an agent, not a role chart.
When the job outgrew the framework#
Skip the library while a single well-built prompt, or a short chain of SDK calls, still holds. Add it when the loop, the tools and the trace would take longer to write than they are worth, then watch for the day the metaphor stops fitting.
Role-and-task modelling stops fitting once the work needs durable execution, fine-grained human approval, or branching that is not a sequence. That boundary is already on the CrewAI page. Recognising it early costs less than discovering it in production, and porting off is a normal outcome rather than a failure.
The same test hits a LangChain chain that reads cleanly at three steps and becomes unreadable at twelve. Keep composition and streaming, and call the provider SDK where the wrapper only adds a stack trace.
Put the bound in your code even while you keep the library. Delegation needs a ceiling, the budget needs to cover the whole run, and a stalled or looping crew has to stop. A failed step has to resume from a checkpoint, not restart the crew, and a person has to see the full arguments on anything that cannot be undone. Those are the controls our harness work already puts around a demo that has to survive traffic.
When those controls will not attach to the library without fighting it, port off. The agent you ship is still the one in the build guide: typed tools, no shell, credentials out of the model's reach, a check on every state change, and a trace a person can read.