Constructing an agent that works in a pocket book takes a day. Getting that very same agent to outlive actual visitors, get well from a crash at 3 am, and never leak another person’s knowledge whereas it runs massive language mannequin (LLM)-generated code is a special job totally, and it is the job most groups underestimate. Solely a really small share of generative AI pilots truly attain manufacturing, and the hole normally is not the mannequin. It is the 5 layers beneath it that no person thinks about till one thing breaks.
This text walks via 5 instruments that shut that hole, one for every layer of the stack: constructing the agent’s logic, executing the code it generates, giving it reminiscence, watching what it does, and working all of it at scale. None of those competes with the others. They sit on prime of each other, and most manufacturing brokers you will encounter in 2026 are working some mixture of all 5.
# 1. LangGraph
A primary agent loop is only a Python whereas loop calling an LLM. That works effective till the loop must department, retry a failed instrument name, pause for a human to approve one thing, or get well after the server it was working on restarts mid-task. At that time, you want one thing that treats agent state as an actual, persevered factor moderately than a variable that disappears the second the method dies.
LangGraph represents an agent as a directed graph as a substitute of a flat chain. Nodes are features, edges join them with non-compulsory conditional routing, and the complete execution is tracked as a sequence of state transitions moderately than a flat message record. Each transition will get checkpointed robotically, which is what makes pause-and-resume, time-travel debugging, and human-in-the-loop approval steps doable with out you constructing that infrastructure your self. Klarna, LinkedIn, Uber, and Replit all run agent workflows on LangGraph, and the framework has change into widespread sufficient in manufacturing settings that its GitHub repository has handed 30,000 stars.
The element value figuring out earlier than you undertake it: the default in-memory checkpointer is okay for improvement, but it surely solely shops state in reminiscence and loses every little thing when the method restarts, which is unacceptable for something actual. Most groups transfer to a Postgres-backed checkpointer the second they go to manufacturing, and that one-line swap is normally the precise level the place a LangGraph undertaking begins behaving like infrastructure as a substitute of a script.
# 2. E2B
The second an agent can write and execute its personal code, you’ve an issue your internet server was by no means constructed to deal with. You possibly can’t run model-generated Python immediately on the identical machine serving your customers, as a result of you don’t have any concept what that code will attempt to do. You want an remoted, disposable setting that may be destroyed the second the duty is finished.
E2B is constructed particularly for this. It focuses on safe sandboxes for AI brokers, specializing in ephemeral code execution with Firecracker microVM isolation, that means every sandbox runs in its personal digital machine with its personal kernel, not only a container sharing the host’s. That is a meaningfully stronger safety boundary than container-based isolation alone. E2B states it’s utilized by 88% of Fortune 100 corporations for frontier agentic workflows, with customers together with Perplexity, Hugging Face, Manus, and Groq.
The tradeoff to know entering into: E2B’s runtime limits are tier-based, capping at one hour on the Interest plan and 24 hours on Professional, so it suits brief, ephemeral execution duties (working a script, testing generated code, a single evaluation job) higher than brokers that want to carry state open for days. In case your agent wants that type of long-lived persistence, that is normally an indication you additionally want the reminiscence layer under, not only a longer-running sandbox.
# 3. Mem0
Each name to an LLM begins from zero except you hand it the related historical past your self. For a single query, that is not an issue. For an agent that is supposed to recollect a consumer’s preferences throughout classes, or choose up a multi-day process the place it left off, a mannequin with no reminiscence is one which quietly forgets every little thing that made it helpful.
Mem0 handles this with out you constructing a customized retrieval pipeline. Throughout a dialog, it extracts the information value maintaining and shops them in a vector database tagged by consumer, session, and agent, then retrieves no matter’s related utilizing a mixture of semantic similarity, key phrase matching, and entity matching earlier than the mannequin responds. The agent seems to recollect the consumer. What’s truly occurring is a focused retrieval step working quietly earlier than each reply, and Mem0 is the commonest drop-in selection for groups that need that with out rolling their very own.
A pure pairing right here is with LangGraph particularly. LangGraph’s personal checkpointers deal with short-term, thread-scoped reminiscence effectively, however they’re designed for dialog continuity and fault tolerance inside a single thread, not for sturdy, cross-thread reminiscence like consumer preferences and information that must persist throughout fully separate classes. That is the hole a devoted reminiscence layer like Mem0 is constructed to fill.
# 4. LangSmith
An agent that fails silently in manufacturing is worse than one which fails loudly, as a result of at the very least the loud failure tells you the place to look. The unglamorous however non-negotiable piece of any manufacturing agent is tracing: a report of each instrument name, each choice, and each remark the agent made alongside the way in which, so when one thing goes mistaken, you are debugging from proof as a substitute of guessing.
LangSmith is constructed for precisely this, and it pairs carefully with LangGraph, although it really works with different frameworks too. It is a industrial agent engineering platform for tracing, debugging, evaluating, and deploying brokers, providing you with a full run-by-run view of what an agent did moderately than simply its closing output. Its free tier consists of 5,000 traces a month with 14-day retention, and the Plus tier runs $39 a seat per thirty days with 10,000 traces, which makes it moderately accessible to strive earlier than committing to it at scale.
What tracing provides you that logging alone does not is the power to replay a particular run and see precisely which step diverged from what you anticipated. That distinction — between figuring out an agent failed and figuring out why — is normally the distinction between a five-minute repair and a multi-day investigation.
# 5. Modal
Even with the logic, sandboxing, reminiscence, and observability sorted, somebody nonetheless has to host all of it, and agent workloads are notoriously bursty: idle for hours, then a sudden spike when visitors hits. Provisioning fastened servers for that sample means both overpaying for idle capability or scrambling when load reveals up.
Modal is a serverless compute platform constructed particularly for this sort of AI workload. It scales from interactive coding brokers to long-running rollouts, spinning up remoted sandboxes that scale to the {hardware} wanted and again to zero when carried out. Modal powers infrastructure for over 10,000 groups, with prospects spanning DoorDash, Anthropic, Meta, and Ramp, and its development has been quick sufficient that Sacra estimated the company hit $300 million in annualized revenue by April 2026, up from roughly $119 million at the end of 2025.
The half that issues most for agent workloads particularly is cold-start time, since no person needs to attend a number of seconds for a sandbox besides earlier than their agent may even begin working. Modal’s GPU reminiscence snapshots can scale back chilly begins by as much as roughly 10x for some workloads, which is the type of element that sounds small till you are working hundreds of brief agent classes a day, and that latency provides up throughout each single one.
# Wrapping Up
None of those 5 instruments is making an attempt to exchange the opposite 4. LangGraph provides your agent’s logic someplace sturdy to dwell, E2B provides it a protected place to run the code it generates, Mem0 provides it a reminiscence that outlasts a single session, LangSmith permits you to see what it truly did, and Modal provides the entire thing someplace to run that scales up and down by itself. The groups that get brokers into manufacturing aren’t those who picked the one finest framework. They’re those who handled every of those as a separate, solvable drawback as a substitute of hoping one instrument would quietly deal with all 5.
When you’re ranging from nothing, the order that tends to work is construct first, sandbox second, and solely add reminiscence and heavier infrastructure as soon as a single agent run is definitely dependable end-to-end. Observability must be wired in from the very first model you ship, not added after the primary incident forces the query.
Shittu Olumide is a software program engineer and technical author captivated with leveraging cutting-edge applied sciences to craft compelling narratives, with a eager eye for element and a knack for simplifying advanced ideas. It’s also possible to discover Shittu on Twitter.
