Uploaded July 2025 | Updated September 2026, 2 weeks ago
đź“Ś Description
What really is an AI agent? Are multi-agent systems worth it — or overhyped? Will one framework rule them all, or will developers stay close to the metal?
🔥 In this super candid, spicy panel, top founders from Crew AI, LlamaIndex, Mastra and Leta dive into the evolving world of agent frameworks. They debate:
What defines an agent vs basic LLM plumbing?
-Are most “agent frameworks” just glorified abstractions?
-Where do protocols like MCP and ADA actually fit (or fail)?
-Is multi-agent orchestration practical, or a demo trap?
-Why agent evaluation (evals) might be overrated — and when it matters.
-What’s next: perpetual agents, no-code builders, dependable systems.
-Perfect for developers, architects, or anyone deep into the future of agentic AI.
⏱ Timestamps
00:02 - Intro: Top agent framework founders on stage (Crew AI, LlamaIndex, Mastra, Leta)
0:26 - The age-old question: What IS an agent? Each founder defines it
1:46 - From LLMs with agency, to closed-loop systems, to control & memory
2:48 - How the term "agent" evolved from 2015 robotics RL to modern LLMs
5:16 - What’s an agent framework? Types in the ecosystem: workflows, orchestration, SDK vs service
6:48 - Why many frameworks are just middleware — or libraries vs servers
8:06 - Spicy take: “LangChainJS sucked — that’s why we built Mastra.”
9:50 - On abstractions: conventions, feature overload, and moving fast
12:50 - Will there be one framework to rule them all, or language-specific stacks?
14:55 - Library vs service trade-offs; easy to swap libraries, harder with hosted state
17:48 - Companies move from playful prototyping to secure production — frameworks commoditize
19:40 - Multi-agent debate: prompt engineering, decomposition, microservices parallels
23:00 - When multi-agent is roleplay vs serious system design
24:35 - Most production use cases today still resemble controlled pipelines, not unconstrained agent swarms
25:40 - MCP vs ADA: why MCP solves tool calling, but ADA is a solution looking for a problem
28:00 - The politics of protocols & vendor lock-in, plus open standards
30:50 - On evals: are they a moat or hype? When customers actually start caring
34:35 - Parallel to research: benchmarking vs method building — vibes first, evals later
36:45 - The next year: stable, perpetual agents; non-techs building agents
#AIagents, #agentframeworks, #CrewAI, #LlamaIndex, #Mastra, #Letta, #MCP
đź“Ś Description
What really is an AI agent? Are multi-agent systems worth it — or overhyped? Will one framework rule them all, or will developers stay close to the metal?
🔥 In this super candid, spicy panel, top founders from Crew AI, LlamaIndex, Mastra and Leta dive into the evolving world of agent frameworks. They debate:
What defines an agent vs basic LLM plumbing?
-Are most “agent frameworks” just glorified abstractions?
-Where do protocols like MCP and ADA actually fit (or fail)?
-Is multi-agent orchestration practical, or a demo trap?
-Why agent evaluation (evals) might be overrated — and when it matters.
-What’s next: perpetual agents, no-code builders, dependable systems.
-Perfect for developers, architects, or anyone deep into the future of agentic AI.
⏱ Timestamps
00:02 - Intro: Top agent framework founders on stage (Crew AI, LlamaIndex, Mastra, Leta)
0:26 - The age-old question: What IS an agent? Each founder defines it
1:46 - From LLMs with agency, to closed-loop systems, to control & memory
2:48 - How the term "agent" evolved from 2015 robotics RL to modern LLMs
5:16 - What’s an agent framework? Types in the ecosystem: workflows, orchestration, SDK vs service
6:48 - Why many frameworks are just middleware — or libraries vs servers
8:06 - Spicy take: “LangChainJS sucked — that’s why we built Mastra.”
9:50 - On abstractions: conventions, feature overload, and moving fast
12:50 - Will there be one framework to rule them all, or language-specific stacks?
14:55 - Library vs service trade-offs; easy to swap libraries, harder with hosted state
17:48 - Companies move from playful prototyping to secure production — frameworks commoditize
19:40 - Multi-agent debate: prompt engineering, decomposition, microservices parallels
23:00 - When multi-agent is roleplay vs serious system design
24:35 - Most production use cases today still resemble controlled pipelines, not unconstrained agent swarms
25:40 - MCP vs ADA: why MCP solves tool calling, but ADA is a solution looking for a problem
28:00 - The politics of protocols & vendor lock-in, plus open standards
30:50 - On evals: are they a moat or hype? When customers actually start caring
34:35 - Parallel to research: benchmarking vs method building — vibes first, evals later
36:45 - The next year: stable, perpetual agents; non-techs building agents
#AIagents, #agentframeworks, #CrewAI, #LlamaIndex, #Mastra, #Letta, #MCP
