Inside the platform
Every Olano agent is a LangChain agent. The model calls, the tool loop, the conversation state, the streaming and the pauses for human approval all run on LangChain's open-source stack. This page sets out which part of that stack does what, where Olano's own work starts, and why we chose to build on it rather than write our own framework.
A lot of AI products are a prompt and a single call to one model provider. An Olano agent is a long-running program: it reads a message from WhatsApp or email, plans, calls tools, waits for a person to approve something, resumes hours later where it stopped, and remembers the thread the next day. Each of those behaviours needs infrastructure underneath it, and for all of them that infrastructure is the LangChain family of libraries.
We build on three of them, each doing a distinct job. Knowing which is which matters mostly when something goes wrong, which is why the rest of this series is about debugging.
From the bottom up: talk to models, run the agent durably, give it the habits of a capable worker.
The model layer. A chat model from Anthropic, OpenAI, Google, Mistral, a local Ollama server or a dozen others is created from a single provider:model string and spoken to through the same message and tool-calling interface. This is what lets an Olano agent switch models without being rebuilt, and lets one deployment run a different model per agent.
The execution layer. The agent's loop - model, tools, model again - runs as a graph, and after every step the state of the conversation is written to a checkpoint keyed by thread. That is why a thread survives a restart, and why an approval can pause a turn indefinitely: LangGraph interrupts the graph, saves where it was, and resumes it when a person decides.
The behaviour layer. On top of LangGraph, Deep Agents gives an agent a to-do list it plans with, file tools over a real workspace, subagents it can hand work to, skills it loads on demand, and automatic summarisation when a conversation outgrows the model's context window. These are the habits that separate an agent that finishes a multi-step job from one that answers a question.
LangChain gives an agent its engine. A business needs the rest of the vehicle.
| Concern | From the LangChain stack | Added by Olano |
|---|---|---|
| Models | A common interface to dozens of providers. | Managed model tiers billed in credits, bring-your-own keys per agent, and a per-agent model picker in the dashboard. |
| Conversation state | Checkpointers that save every step of a thread. | Production storage for those checkpoints, encryption at rest, and deletes that remove a thread from every store that holds it. |
| Human approval | Interrupts that pause and resume a graph. | Trust levels, approval categories, an Approvals inbox, and rules that keep the riskiest actions waiting even in automatic modes. |
| Tools | Tool calling, plus adapters for MCP servers. | A catalogue of connectors, sign-in links for each service, a credential vault scoped per agent, and repair when a service revokes access. |
| Reaching people | Nothing: the libraries do not talk to WhatsApp. | WhatsApp, Telegram, Slack, email and the web dashboard, each with its own delivery rules. |
| Running unattended | Nothing: a graph runs when it is invoked. | Schedules, heartbeats, issue polling, and Cortex, the background cycles that improve the system with your approval. |
| Operations | Libraries you install. | Managed deployments, isolated sandboxes, upgrades tested against the pinned library versions, and diagnostics in the dashboard. |
Breadth we could not match alone. The integration packages cover far more model providers, vector stores and tool protocols than a small team could maintain. When a new model ships, the provider's LangChain package usually supports it within days, and an Olano agent can use it as soon as we have tested the upgrade.
A runtime built for state, not just calls. Durable checkpoints, interrupts and resumable streams are the hard parts of running agents for a business, and LangGraph treats them as first-class. Writing our own would have meant rediscovering the same edge cases more slowly.
Less lock-in for you. An agent's behaviour is described in standard terms - a model string, tools, skills as plain Markdown files, MCP servers - that the wider ecosystem understands. We wrote about why that matters in AI agent platform lock-in.
The honest trade-off. A fast-moving upstream means upgrades need care. We pin the Deep Agents version exactly, raise the LangChain version floors together with it rather than one at a time, and run the whole test suite, including live sandbox tests, before a new combination reaches a customer deployment. Most of what we learned doing that is in the four debugging guides below.
You do not need to know any of this to run an Olano agent: the dashboard, the chat and the approvals inbox are the interface. It shows up indirectly. You can change an agent's model without losing its conversations. A thread picks up where it left off after the deployment restarts. An approval really does pause the work rather than asking after the fact. And when we tell you a behaviour comes from the underlying library, you can read the library's own documentation and check.
What running LangChain agents in production taught us, written up for anyone building on the same stack.
The toolbox: debug mode, LangSmith traces, stream modes, state history, loop limits and fake models for reproducing a bug offline.
Missing thread ids, memory lost on restart, "bound to a different event loop", Postgres setup, encryption and deleting a thread properly.
Why tokens from a model called inside a tool leak into the reply, why two of them garble it, and why a stream that has started must never be retried.
Content blocks one provider writes and another rejects, models with no known context window, and parameters that change name between providers.
No. Olano is an independent company based in Singapore. The Olano platform is built on LangChain's open-source libraries - LangChain, LangGraph and Deep Agents - and adds channels, approvals, a credential vault, scheduling, Cortex and managed hosting on top. Olano is not affiliated with or endorsed by LangChain, Inc.
No. Agents are created and configured in the Olano dashboard or by asking an agent in chat. LangChain matters to you only indirectly: it is why an agent can change models without losing its threads, and why an approval genuinely pauses work until someone decides.
Olano's managed model tiers, which run on credits with no provider key, or your own key for any provider with a LangChain integration that Olano supports, including Anthropic, OpenAI, Google, Mistral, Groq, DeepSeek, xAI, AWS Bedrock, OpenRouter and local Ollama models. Each agent can use a different model.
The libraries run inside your own deployment and do not send anything to LangChain, Inc. LangSmith, LangChain's hosted tracing service, is a separate product that Olano does not switch on. Conversations go only to the model provider the agent is configured to use.
Conservatively. The Deep Agents version is pinned exactly, the LangChain package floors are raised together with it, and every new combination passes the full test suite, including tests against real sandboxes, before it is released to customer deployments.
The rest of the LangChain series, and where to go next.
Debug mode, LangSmith traces, stream modes, state history, loop limits and fake models.
Thread ids, lost memory after a restart, event-loop errors, Postgres setup, encryption and deletes.
Leaked tokens from nested model calls, garbled replies, missing paragraph breaks and dropped connections.
Content blocks one provider writes and another rejects, missing context windows, and retries that should not happen.
One system, the right model for every task - the product view of the model layer.
Channels, approvals, memory and Cortex, the parts Olano builds on top of the stack.
Checked 13 October 2026 against LangChain 1.4, langchain-core 1.6, LangGraph 1.2 and Deep Agents 0.7. Olano is an independent company that builds on LangChain's open-source libraries; it is not affiliated with or endorsed by LangChain, Inc. Library behaviour changes between releases, and the LangChain documentation is the primary source. If we have described something wrongly, write to support@olano.ai and we will correct this page.
Whether you want agents running on LangChain without operating them yourself, or a second opinion on an agent you have built, tell us what you are working on.
