How agents will unlock the $500B promise of AI - Donald Hruska, Retool

Introduction and Retool's Entry into Agentic AI 00:00

• Retool, known for simplifying internal application development, has launched Retool agents, expanding into agentic AI.
• Enterprises have invested around $500 billion in AI infrastructure, yet most only use simple chatbots or code generators.
• The landscape is shifting to allow enterprises to build AI agents with guardrails that connect to real production systems.

AI Market Growth and Impact on Coding Workflows 01:02

• Anthropic’s annualized revenue grew from $1B in December to $3B at end of May, a 3x increase in 5 months.
• OpenAI is projected to reach $12B in revenue by end of 2025, more than tripling their previous year's figures.
• These dramatic increases are driven primarily by enterprise AI spending.
• Coding workflows are being transformed by AI tools such as Cursor and Windsurf, raising developer productivity.
• OpenRouter’s top apps are dominated by code generation, highlighting this major enterprise use case.
• LLM coding capabilities keep improving; GPT-4.1’s score on SWEBench is up 21 points over GPT-4.0; Gemini 2.5 Pro is up another 9 points over GPT-4.1.
• Developers favor new models like Gemini 2.5 Pro for coding tasks, with widespread positive feedback.

Vibe Coding and Shift Toward Agentic AI 03:00

• Vibe coding, likened to “the punk rock of software,” is democratizing the ability for anyone with an idea to create applications.
• Unlike basic code completion, vibe coding leverages agents that autonomously think, act, and write code based on high-level instructions.
• Code is a natural domain for agents due to its semantic clarity and testability; the question is whether this agentic pattern can expand to other business problems via general-purpose agents.

Basics and Challenges of Building Agents 04:07

• Building a basic agent is technically easy; an agent is essentially an LLM in an execution loop, able to read, decide, call tools, and self-verify using frameworks like ReAct.
• Real production challenges go beyond code: security, access control, integration, audit logs, compliance, internationalization, and more can't be solved by "vibe coding" alone.
• Risks exist with AI-generated code, including vulnerabilities and hallucinations; real-world issues have resulted from insufficiently vetted AI logic.
• Important safeguards include security controls, limiting agent access, managing costs, and using evaluations (evals) to ensure as much determinism as possible.

Frameworks and Strategies for Deploying Agents 07:26

• Four approaches to building and deploying agents:
  • Build from scratch: Maximum control, high lift, custom integrations, suitable for core or sensitive domains.
  • Use frameworks (e.g., LangGraph): Medium lift, flexible, still significant engineering.
  • Use agent platforms (e.g., Retool agents): Low lift, opinionated defaults, fast to production, platform-dependent, strong for a broad range of business use cases, built-in connectors, observability, and compliance features.
  • Use verticalized agents: Do one thing well, minimal flexibility, ideal for specific tasks but not adaptable elsewhere.
• The choice is an engineering trade-off: build for core, competitive, or regulated needs; buy for commodity workflows needing speed.
• Risk assessments and feature comparisons (e.g., connectors, permissioning, compliance, audit, and observability) are crucial for managed platform decision-making.

The Build vs. Buy Decision in Practice 10:41

• Analogy to broader software decisions (e.g., Stripe builds core billing by hand, uses platforms for non-core needs).
• As companies scale, they may shift commodity or high-volume work to managed agent platforms, retaining hand-built solutions for areas of differentiation.
• Example: Cursor would hand-build core product logic, but might use agent platforms for peripheral operations.

Real-world Impacts and Business Value of Agents 12:03

• Companies like AWS are leveraging agents to automate routine business processes.
• ClickUp saved $200,000+ in vendor costs and avoided large headcount increases by building on Retool.
• Descript saves hundreds of hours weekly through 50 internal apps built with agentic tools.
• Retool reports automating over 100 million hours of work cumulatively for customers.
• Automating routine work frees up humans for creative and strategic efforts, drawing a parallel with the democratizing impact of the printing press.
• Mass adoption of agents is expected to enhance team capabilities and drive global productivity growth.

Industry Trends: Costs and Adoption 13:05

• AI inference costs have dropped dramatically: 99.7% reduction in per-token cost from 2022 to 2024.
• Retool’s cheapest agent now costs $3/hour, with expectations for costs to decrease further.
• Google searches for “AI agents” increased by 11x over 16 months, indicating surging interest and adoption.

Closing Thoughts and Q&A 14:09

• There’s no universal solution for automating all business processes; focus should be on using agents where they create the most leverage.
• Retool’s internal philosophy is to build as much as possible using their own platform, building custom solutions only when necessary.
• On-premises support for Retool agents is being added, including for airgapped environments, in response to customer needs for more secure deployments.