In Data Science Connect's Plug & Play Q3 webinar, Precisely showed an agent that checks data quality scores before it acts and repairs failing data through governed pipelines. Coder showed coding agents running in self-hosted, sandboxed workspaces with budgets, session logs and human-owned pull requests.
Key takeaways
- Agents remove the human checkpoint. Martini said bad data stops being a quality problem and becomes an operational one when agents act on it.
- Gate every agent call. Precisely's trust gate checks quality scores against a use-case threshold and blocks analysis that falls short.
- Fix, enrich, rescore. In the demo, the agent ran a governed cleansing pipeline, enriched the records and rescored them before answering.
- Move coding agents off laptops. Coder runs agents in self-hosted, sandboxed workspaces with spending budgets and full session logs.
- A human owns every change. Coder's agents open draft pull requests under the name of the person who assigned the task.
Two vendors ran live demos in Data Science Connect's Plug & Play: Live AI Workflows in Action Q3 webinar, a hands-on session sponsored by Precisely and Coder. Precisely showed an agent that checks data quality before it acts. Coder showed how coding agents can run in governed, self-hosted workspaces instead of on laptops.
The presenters were Anna Martini, Product Marketing Manager at Precisely; Traci Hudgin, Product Specialist at Precisely; and Ben Potter, Product at Coder. Jenna Holland of Data Science Connect hosted. Here's what each demonstrated.
Why does agentic AI raise the stakes on data quality?
Martini said AI programs run into three walls: trust, data sprawl across warehouses, lakehouses, mainframes and SAP, and the difficulty of proving outcomes. She focused on trust. A model working from bad data, she said, "is going to hand you a very confident and very wrong answer."
In traditional analytics, a person who sees a number that looks off goes and investigates. Agents remove that checkpoint and act on whatever they are handed, Martini said, so errors move through the workflow unnoticed. As a result, "bad data goes from being a quality problem to a much riskier operational one."
She showed the same model giving two different answers to "what is our revenue." The difference was the data underneath: governed, clean data with semantics versus data with no quality checks, ownership or audit trail.
What is an AI trust gate?
Martini placed a data integrity layer between metadata and lineage and the agents that use the data. It covers quality scoring, governance, enrichment and certifying data as fit for AI use, and agents can query those trust signals when they decide what to do.
In the pattern she described, every agent task passes through a trust gate. If the data clears the threshold set for that use case, the analysis proceeds. If not, the agent explains the business risk in plain language, cleanses and enriches the data through a governed pipeline, then rescores it. Certification cannot be a one-time step, she said: "you need a trust gate that is continuously firing on every single agent call."
Precisely: How does the trust gate work in a live workflow?
Hudgin used VS Code with GitHub Copilot chat, Precisely's MCP server and the Snowflake MCP server, with raw account data in Snowflake. Her prompt asked whether to expand headcount in Southeast markets based on existing accounts.
The agent applied a 90% quality threshold, found the data scored 77% and refused to estimate the market opportunity. It then surfaced an existing data quality pipeline that standardized inconsistent state values, using address fields to locate records. Hudgin ran the pipeline from the MCP server, and the rescored data passed the threshold.
She then enriched the records with neighborhood demographics, walkability, growth trends and tax jurisdictions. The agent handled API pagination errors on its own and then ran the verified steps across the full data set. Its final output recommended Georgia, then North Carolina.
Asked why an agent should go through Precisely rather than straight to enterprise data, Hudgin pointed to business definitions. Connected straight to raw data, she said, "it's going to make an assumption based on the values that it sees and probably the column headers."
Coder: Why move coding agents off the laptop?
Potter said more people are writing software with coding agents, including business users such as finance staff building internal workflows. In large regulated organizations, he said, it is hard to get those agents onto laptops and harder still to connect them safely to company data. "We've outgrown using laptops for the developer environment," he said.
Coder is self-hosted and can run air-gapped on an organization's own cloud or on-premises infrastructure. Users sign in through single sign-on and pick templates that a platform team defines for each use case, such as software engineering or data science. Potter launched a workspace with VS Code in the browser, a cloned code base and Claude Code already set up. He said the agent's autonomous mode is safer there than on a laptop because the workspace is sandboxed.
He also showed per-user AI spending budgets tied to group identity, and a session log that records each prompt and every tool call and command an agent ran.
How do you keep a human accountable for agent-written code?
Potter assigned a Git issue to an agent bot, which spun up a workspace, worked the task and opened a draft pull request under his name. The agent had limited internet access and no permission to touch production. "I assign the agent to do this. This is my work," he said. The developer then reviews the draft before a peer does.
Asked what is risky on a laptop, he cited what AI researcher Simon Willison calls the "lethal trifecta": access to private data, the ability to communicate externally and exposure to untrusted content. The risk grows, he said, as people with less security background start running agents.
On whether review slows teams down, Potter said Coder has not seen a large review bottleneck. "The cost of generating code has become really cheap," he said, which makes clear priorities and good judgment more important.
See more live AI demos
Plug & Play is one of four series in Data Science Connect's 2027 webinar lineup, alongside the AI × Enterprise Panel Series, AI Demo Days and Agentic AI Showcases. The schedule includes 24 free live webinars, with registration coming soon. See the webinar schedule for upcoming sessions.
Frequently asked questions
What is a trust gate for AI agents?
A check that runs before an agent acts. If the data's quality score clears the threshold set for that use case, the analysis proceeds; if not, the agent explains the risk, cleanses and enriches the data, and rescores it.
Why connect an AI agent to data through an MCP server instead of directly?
Hudgin said an agent on raw tables guesses from values and column headers, while Precisely's MCP server supplies business definitions, quality scores and semantic models for more consistent answers.
Why run coding agents in remote workspaces instead of on laptops?
Potter said sandboxed, self-hosted workspaces let organizations limit agent permissions and internet access, set budgets and audit every tool call, which is harder across many laptops.
What is the lethal trifecta for AI agents?
As Potter described it, giving an agent access to private data, the ability to communicate externally and exposure to untrusted content at the same time.
Watch: Plug & Play: Live AI Workflows with Precisely & Coder | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
