SUMM

The talk addresses the debate between using agents versus workflows in AI, asking why not use both.
References a recent debate in the AI community, especially on Twitter and in blog posts from Anthropic and OpenAI.
Controversy arose around OpenAI's stance that seemed dismissive of workflow approaches, stirring responses in the community.
Highlights the influential role of large model providers in shaping discourse due to their elevated position.

Warns against "that guy" mentality, where some believe there is only one right way to develop systems, especially when coming from dominant companies.
Notes historical parallels from past web development debates, especially with technologies pushed by major companies.
Criticizes “graph node and edge” terminal APIs in workflow and agent frameworks, arguing they reduce readability and accessibility for teams.
Advocates for more intuitive, readable, and linear workflow syntax over graph-based representations, favoring patterns that don't require understanding graph theory.

Draws from architectural design pattern literature, noting its influence on software engineering.
Points out there is still a lack of standardized language and patterns for composing agents and workflows in AI systems.
Defines agents as interactive, turn-based components (like turn-based games), while workflows are sequences of dependent steps (like tech tree progression in strategy games).
Emphasizes that tracking dependencies and handling non-determinism are especially important in AI systems.

Discusses trade-offs between power and control in system design—sometimes you need more control (workflows), sometimes more power (agents).
Suggests breaking down complex or unreliable processes into more structured, manageable workflow steps to increase reliability.
Recommends diagramming and explaining architectures to peers as a way to uncover better, more creative solutions.

Explains that agents and workflows can be composed in multiple ways: agents can use other agents as tools, agents can be workflow steps, workflows can function as tools for agents, etc.
Describes the agent supervisor model where a supervising agent orchestrates other agents performing specialized tasks.
Notes that the real benefits come from combining these patterns creatively.

Dynamic tool injection is discussed: limit the number of tools an agent accesses simultaneously to improve performance.
Recommends thoughtful assignment of tools to agents, especially as tasks get more complex.
Encourages community members to share practical experiences and adapt theory based on what works in practice, as the field is evolving quickly.
Asserts that effective solutions may be ahead of formal theoretical understanding in this rapidly developing area.

Responds to a question about agents using many tools: if it works in practice, continue with that approach even if theory suggests otherwise.
Emphasizes that practice is evolving faster than theory in the field of agentic workflows.
Provides information for connecting after the talk and encourages attendees to pick up a copy of the speaker’s book.

Agents vs Workflows: Why Not Both? — Sam Bhagwat, Mastra.ai