EP 51: Building AI Agents That Actually Work in Production | Shawn Kercher, CEO @HummingAgent.ai

Summary

This Data Science With Sam episode has Sam interview Shawn Kercher of HummingAgent about building private, domain-specific agents around real business workflows rather than packaging a general model behind a new interface. Its durable synthesis is that production value comes from proprietary context, legacy-system integration, workflow design, and controlled authority: Production Agent Discovery should begin with real data, Progressive Agent Autonomy should expand only after evidence and feedback, and Model-Portable Agent Architecture should keep the application testable as models change.

Key Claims

• HummingAgent begins with paid, fixed-scope discovery to understand a client’s systems, APIs, operational rules, employee knowledge, and real data before implementation.
• Production Agent Discovery should use a copy of the actual production data because clean demonstration data hides incomplete CRM records, missing fields, exceptions, and undocumented practices.
• A defensible agent product is the surrounding workflow and integration layer: company-specific rules, proprietary data, legacy software, security boundaries, and user trust are harder for a foundation-model provider to reproduce than a generic connector.
• Progressive Agent Autonomy starts with a narrow sandboxed capability, keeps people responsible for consequential actions such as sending email, and widens access only after testing and feedback.
• Human-in-the-Loop Agent Governance remains necessary because supervision, subject-matter expertise, and approval can catch contextual errors that a broadly authorized agent may turn into operational or security harm.
• Model-Portable Agent Architecture avoids dependence on one provider, but changing models creates a continuing quality-assurance and evaluation burden because models behave differently across workflows.
• The episode presents narrow vertical systems as stronger near-term candidates than broadly autonomous agents and predicts pressure on per-seat SaaS pricing as agents perform more software-mediated work.
• A source-reported client case says AI search reduced research across 44 years of PDF documents from hours to minutes, but the supplied episode note provides no architecture, error rate, cost, or independent evaluation.

Key Quotes

No reliable verbatim quotations are retained. The supplied document is a structured episode summary rather than a transcript, and the episode’s “wiring” metaphor is treated as a paraphrased claim.

Connections

• Data Science With Sam, Sam (Data Science With Sam), Shawn Kercher, and HummingAgent - show, host, guest, and company context.
• Production Agent Discovery and AI Data Readiness - discovery, real-data inspection, and undocumented operational knowledge before implementation.
• Progressive Agent Autonomy, Human-in-the-Loop Agent Governance, and Agent Permission Boundaries - sandboxing, review, and graduated authority.
• Model-Portable Agent Architecture, Model Routing Cost Control, and Agent Reliability Verification - provider portability, workflow-specific model choice, and regression testing.
• Domain Know-How Moat, Organizational Context, and Regulated Enterprise AI Deployment / 受监管企业AI部署 - proprietary context, legacy systems, and regulated private deployment.
• Outcome-Based AI Pricing - adjacent implication of agents weakening conventional per-seat software pricing.

Contradictions

• The source metadata calls the guest “Shawn Kercher” and a CEO, while the body calls him “Sean Kerscher” and HummingAgent’s co-founder and CTO. This ingest uses the title’s spelling for the canonical identity but leaves name, title, and role unresolved.
• No settled contradiction with the existing wiki is adopted. The episode reinforces existing permission, human-review, data-readiness, and domain-knowledge claims while specializing them to small- and medium-business agent deployments.
• The 44-year archive result, private-data assurances, productivity gains, workforce effects, client examples, and forecasts about SaaS pricing and narrow-agent adoption remain participant-reported or source-scoped.