In a hands-on Plug & Play webinar, Teradata built a LangChain agent that checks loyalty discount eligibility using a vector store and a governed SQL tool. Coder explained what its experiments with agent harnesses, sub-agents and context management did and didn't improve.
Key takeaways
- An LLM alone isn't an agent. An agent combines a goal, tools, an LLM and a runtime loop that reasons toward the goal.
- Govern what agents can query. Teradata's demo agent used only approved MCP tools instead of generating its own SQL.
- Explainable answers cite sources. The agent returned its decision with the policy clauses and SQL it used.
- Agents need sandboxes like developers do. Coder applies the same workspace controls to agents as to human engineers.
- Optimization is trial and error. Coder's cheaper output-filtering experiment ran faster but cut results; sub-agents kept the main context small.
Two engineers built and ran AI agents live in Data Science Connect's Plug & Play: Live AI Workflows in Action Q1 webinar, a hands-on session sponsored by Teradata and Coder. Teradata showed how to build an agent that reasons over policy documents and database tables through governed MCP tools. Coder showed what it learned from experimenting with agent harnesses, sub-agents and context management.
The presenters were Janeth Graziani, Developer Advocate at Teradata, and Michael, an engineering manager at Coder. Here's what each demonstrated.
What is an AI agent, and what is it made of?
Graziani started with a definition. An agent, she said, is a system that is given a goal, a set of tools and an LLM, and uses the LLM to reason toward the goal with those tools. By that standard, a basic browser chatbot isn't an agent, and neither is a model on its own. "The LLM is not an agent," she said.
She broke an agent into four parts: inputs, a runtime, the LLM and tools. The developer's system prompt sets the boundaries, and the user's prompt is the task within them. The runtime keeps state and message history, calls the LLM repeatedly, executes tool calls and appends results until the LLM decides it has enough to answer. Frameworks such as LangChain and LangGraph abstract much of that loop, and Teradata integrates with them.
Teradata: How do you build an agent that reasons over documents and tables?
Graziani's lab, which attendees could follow in Teradata's free hosted environment, posed a finance team's question: which customers qualify for a loyalty discount? The eligibility policy lives in unstructured documents that change. The customer facts live in structured tables. The agent has to combine both and produce an explainable, audit-ready answer.
"This is very different than your typical text to SQL agent," she said. "The agent doesn't generate SQL." Instead, it works only through approved MCP tools.
In a Jupyter notebook, she loaded sample policy clauses, including a 12-month tenure requirement, hard exclusions, regional exceptions and a rule that government accounts always need manual review, into a Teradata enterprise vector store with Gemini embeddings. She then declared a custom SQL tool in a YAML file that returns only the business attributes needed for an eligibility check, and registered it under a finance analyst profile together with the vector store tools. The open-source Teradata MCP server exposed those tools to a LangChain agent.
Asked whether a sample customer, Acme Co, was eligible, the agent ran a similarity search for the policy clauses, called the SQL tool, and returned a decision with the supporting clauses and the exact SQL executed. Graziani said Teradata also offers tools for the later deploy and monitor stages of the agent lifecycle, which she said she hopes to cover in a future session.
Coder: Why do AI agents need governed development environments?
Michael said Coder began as a cloud development environment: self-hosted workspaces that platform and security teams define and control, which developers reach from their IDEs and terminals. When AI agents arrived, Coder saw that "the same principles that we have been applying to human builders also apply to agent builders," he said. An agent is compute that wants tools, and its sandbox can be defined like a workspace, with limits on internet access and source code.
He argued that cost will matter once the hype fades, as it did with cloud computing, but said efficiency means the best quality per unit spent, not the lowest spend. A cheaper model may answer faster but need more human correction.
What techniques make agent harnesses more efficient?
Michael described the agent harness, the control layer around model calls, tools and context, and several experiments Coder ran:
- Programmatic tool calling: having the agent write one script that, for example, runs a query for five parameters, instead of making five separate tool calls.
- Context management: windowing, summarizing and compacting history, since some evaluations suggest models start to drift once 50% to 60% of the context window is used.
- Sub-agents: delegating tasks to smaller, cheaper models that return only relevant results. "They do the grunt work," he said, "and they only come back with the relevant pieces of information."
Coder tested these with an open-source coding agent multiplexer it built, against nine randomly chosen Terminal Bench tasks. One experiment, using small models to filter tool output before it reached the main agent, was cheaper and faster but produced worse results because the main agent lost context. Michael said Coder's own harness with an OpenAI model outperformed OpenAI's Codex harness on those tasks, at higher cost than its Anthropic baseline. No one, he said, can give "a super well-described approach that will solve all of your problems."
In a demo, the multiplexer created a plan to add a bulk cancel option to Coder's dashboard, then spun up sub-agents. The main agent's context held about 26,000 tokens, even though much more had been processed. He also showed Coder Agents, in early access, a simpler interface that runs agents on Coder's own infrastructure using the same workspace templates as its engineers. An open-source library called Portable Desktop lets an agent use a Linux desktop. In the demo, the agent opened Chrome and confirmed the new button appeared.
See the next Plug & Play session
Plug & Play is one of four series in Data Science Connect's 2027 lineup of 24 free live webinars, alongside the AI × Enterprise Panel Series, AI Demo Days and Agentic AI Showcases. See the webinar schedule for upcoming hands-on sessions.
Frequently asked questions
What is the difference between an LLM and an AI agent?
An LLM is a generative model. An agent, as Graziani defined it, is a system that uses an LLM to reason toward a goal with a set of tools, inside a runtime loop.
What is an agent harness?
The control layer around an LLM that decides how the model is called, which tools it gets, what context it receives and how its behavior is controlled, as Coder's presenter described it.
How can an AI agent use enterprise data safely?
In Teradata's demo, the agent accessed data only through predefined MCP tools registered to a role, including a custom SQL tool limited to the needed attributes.
Why use sub-agents?
Sub-agents, often running smaller models, handle exploration and return only relevant results, which keeps the main agent's context window small.
Watch: Plug & Play Live AI Workflows in Action Q1 | Hands-On Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
