Most AI agent prototypes never become trusted production systems. At COLLIDE Data + AI Conference 2026, LangChain Deployed Engineer Christine Manegan said teams that succeed own their intelligence: the models, harness, evals and governance, and explained how an agent harness manages context, sub-agents and human-in-the-loop controls.
Key takeaways
- Own your intelligence. Manegan said production teams control their models, scaffolding, context, evals, observability and governance.
- Agents are model plus harness. The harness of tools, context, sub-agents, prompts and skills is what lets a model act.
- Agents aren't regular software. Non-deterministic outputs make continuous monitoring ongoing feature work.
- Deterministic middleware adds control. Hooks around tool and model calls enforce business rules, PII redaction and human approval.
- Production needs infrastructure. Sandboxes, model gateways with cost policies, and horizontal scaling keep mission-critical agents running.
Many enterprises have one or hundreds of AI agent prototypes. Far fewer have agents they trust with real data and real users. At COLLIDE Data + AI Conference 2026 in Atlanta, Christine Manegan, a Deployed Engineer at LangChain who works with teams shipping agents into production, explained the architecture she sees behind the ones that make it.
Prototypes are not the hard part, Manegan said in her session, "The Architecture Behind Mission-Critical Agents." The hard part is "getting those agents into an actual production state" where they are trustworthy and reliable.
What do production AI agents have in common?
The companies that get agents from prototype to production share one trait, according to Manegan: they own their intelligence. She broke that into three parts.
- Open systems. Teams should be model agnostic, able to switch models or providers if token costs climb or a model starts behaving badly. They also need to own the scaffolding around the model, including the system prompt, sub-agents, skills and tools, and the agent's context and memory, so conflicting information can be found and removed rather than hidden in a black box.
- The improvement layer. Teams should define their own evals, because what counts as good differs by use case. A financial analyst agent needs different metrics than an email assistant, she said. They also need to own the observability data, the traces of what the agent actually does in production.
- Governance. That covers uptime for the infrastructure where agents run, cost controls, and an authorization model for which tools an agent can reach and who can interact with it.
How is the agent development life cycle different from software development?
The agent life cycle of build, test, deploy and monitor resembles the software development life cycle, Manegan said, with one key difference: agent outputs are non-deterministic. Traditional code can be unit tested against expected outputs. An agent reasons its own way through each step, so the same input can produce different results.
An agent is "only as good as the context it has around it," she said, and that context changes as data and tools change. Monitoring how agents behave and how people use them is therefore ongoing feature work, not just bug fixing.
What is an agent harness?
"An agent is comprised of a model and the harness around it," Manegan said. The model, usually an LLM, does the reasoning. The harness is the tools, context, sub-agents, prompts and skills that let it act in an environment.
"At the core of every agent is a model that's calling tools in a loop," she said, calling tools, getting results and deciding again until the task is done.
Telling an off-the-shelf agent to "make no mistakes" is not a production strategy, Manegan said. A custom harness lets a team connect the agent to its own data, manage context over long runs, run tasks in parallel, add human steering and improve over time. Context needs vary widely: an agent on a trading desk needs different information, faster, than one doing long, deep research.
How does LangChain's Deep Agents harness work?
Manegan presented Deep Agents, LangChain's open-source harness, which she said sits on top of the company's LangGraph and LangChain frameworks. She described three properties.
Provider agnostic. Any frontier or open-source model can be used. Manegan said LangChain has seen strong coding-benchmark results by tuning the harness rather than the model. She also described a common pattern: build with a frontier model, turn its behavior into a golden data set, fine-tune a smaller and cheaper model on it, and plug that model into the same harness.
Customizable. Agents get file system tools such as ls, glob and grep, which LLMs use well. The file system can be virtual and backed by a database that holds long-term memory. Context management includes skills, summarization, offloading long tool outputs to files, and prompt caching. Sub-agents handle context-isolated tasks, such as reading a long article and passing back only a summary.
Steerable. Middleware lets teams insert deterministic steps around tool calls or model calls, such as checking that a report has the required sections or redacting PII before data reaches the model. Manegan said human-in-the-loop approvals are one form of middleware, and "human in the loop is always going to matter to some extent" for mission-critical agents.
What infrastructure do mission-critical agents need?
Production is "a beast in and of itself," Manegan said. LangChain's managed Deep Agents offering, she said, adds:
- Sandboxes for secure code execution, which she said agents need for presentations, PDF processing, calculations and spreadsheets, not just coding
- A gateway between the agent and the model to redact PII and enforce cost policies, avoiding "runaway agents that are driving up to millions of tokens overnight"
- Tool servers that control authentication
- A central context hub for skills and prompts, so domain experts can stay close to an agent's context without being developers
- Streaming, durability and checkpointing, and horizontal scaling
Explore agentic AI at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Between conferences, Data Science Connect's free Agentic AI Showcases are among 24 live webinars on the 2027 schedule.
Frequently asked questions
What is an AI agent harness?
As Manegan described it, the tools, context, sub-agents, prompts and skills around a model that let it interact with an environment and take the right actions.
How is the agent development life cycle different from the SDLC?
Agent outputs are non-deterministic, so unit tests aren't enough. Teams must continuously monitor agents, evaluate them and update them as data and tools change.
What is LangChain Deep Agents?
An open-source agent harness from LangChain, built on LangGraph and LangChain, that provides file system tools, context management, sub-agents and middleware.
Why do production AI agents need human in the loop?
Manegan said human approval before sensitive actions, such as database writes or sending email, will always matter to some extent for mission-critical agents.
Watch: The Architecture Behind Mission-Critical Agents on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
