StarZero All shows Try the podcast clipper →

Jul 5, 2025 · 27m · latent-space

⚡️Anthropic vs Cognition on Multi-Agents: A Breakdown with Dylan Davis

Dylan Davis · 16m spoken Shawn Wang · 6m spoken
0:00 / 0:00
▶ Watch on YouTube →

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

Swix and Dylan Davis break down the competing AI agent philosophies of Anthropic and Cognition, comparing multi-agent parallel research architectures with single-agent sequential coding systems to provide an engineering decision framework.

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 29.6% of the talking time here. How this is scored →

The hosts as informed peer 6.3 Guest teaching 2.8 Guest disagreement 1.0 The hosts pushing back 3.0
05100:0010:0020:001:39–7:03 · The hosts as informed peer 5/10 Deconstructing Anthropic's Multi-Agent Architecture for Deep Research Dylan walks through Anthropic's deep research multi-agent architecture from his visual breakdown. Swyx lightly questions Dylan's skepticism regarding Anthropic's parallel execution claims and points out the underlying communication complexity that diagrams omit.7:03–14:02 · The hosts as informed peer 7/10 Performance Trade-Offs, Token Costs, and Evaluation Frameworks Dylan explains Anthropic's 15x token trade-off and five evaluation dimensions. Swyx demonstrates deep domain knowledge by explaining LLM judge consolidation mechanics, context engineering, and the safety contrast between read-only research and write-enabled coding agents.14:04–19:52 · The hosts as informed peer 6/10 Why Cognition Favors Single-Agent Sequential Architectures for Coding Dylan demonstrates Cognition's sequential single-agent approach using a Flappy Bird example. Swyx questions whether the scenario is contrived, catches Dylan off guard by verifying the exact text in Cognition's post, and contextualizes Cognition's view against OpenAI research.19:53–24:15 · The hosts as informed peer 7/10 A Decision Framework for Selecting Agent Architectures Dylan shares decision heuristics and Claude artifacts for agent selection. Swyx offers an authoritative counter-perspective on token economics, arguing that optimizing for cost is a trap because the price of intelligence drops 100x annually.1:39–7:03 · Guest teaching 3/10 Deconstructing Anthropic's Multi-Agent Architecture for Deep Research Dylan walks through Anthropic's deep research multi-agent architecture from his visual breakdown. Swyx lightly questions Dylan's skepticism regarding Anthropic's parallel execution claims and points out the underlying communication complexity that diagrams omit.7:03–14:02 · Guest teaching 3/10 Performance Trade-Offs, Token Costs, and Evaluation Frameworks Dylan explains Anthropic's 15x token trade-off and five evaluation dimensions. Swyx demonstrates deep domain knowledge by explaining LLM judge consolidation mechanics, context engineering, and the safety contrast between read-only research and write-enabled coding agents.14:04–19:52 · Guest teaching 2/10 Why Cognition Favors Single-Agent Sequential Architectures for Coding Dylan demonstrates Cognition's sequential single-agent approach using a Flappy Bird example. Swyx questions whether the scenario is contrived, catches Dylan off guard by verifying the exact text in Cognition's post, and contextualizes Cognition's view against OpenAI research.19:53–24:15 · Guest teaching 3/10 A Decision Framework for Selecting Agent Architectures Dylan shares decision heuristics and Claude artifacts for agent selection. Swyx offers an authoritative counter-perspective on token economics, arguing that optimizing for cost is a trap because the price of intelligence drops 100x annually.1:39–7:03 · Guest disagreement 1/10 Deconstructing Anthropic's Multi-Agent Architecture for Deep Research Dylan walks through Anthropic's deep research multi-agent architecture from his visual breakdown. Swyx lightly questions Dylan's skepticism regarding Anthropic's parallel execution claims and points out the underlying communication complexity that diagrams omit.7:03–14:02 · Guest disagreement 1/10 Performance Trade-Offs, Token Costs, and Evaluation Frameworks Dylan explains Anthropic's 15x token trade-off and five evaluation dimensions. Swyx demonstrates deep domain knowledge by explaining LLM judge consolidation mechanics, context engineering, and the safety contrast between read-only research and write-enabled coding agents.14:04–19:52 · Guest disagreement 1/10 Why Cognition Favors Single-Agent Sequential Architectures for Coding Dylan demonstrates Cognition's sequential single-agent approach using a Flappy Bird example. Swyx questions whether the scenario is contrived, catches Dylan off guard by verifying the exact text in Cognition's post, and contextualizes Cognition's view against OpenAI research.19:53–24:15 · Guest disagreement 1/10 A Decision Framework for Selecting Agent Architectures Dylan shares decision heuristics and Claude artifacts for agent selection. Swyx offers an authoritative counter-perspective on token economics, arguing that optimizing for cost is a trap because the price of intelligence drops 100x annually.1:39–7:03 · The hosts pushing back 2/10 Deconstructing Anthropic's Multi-Agent Architecture for Deep Research Dylan walks through Anthropic's deep research multi-agent architecture from his visual breakdown. Swyx lightly questions Dylan's skepticism regarding Anthropic's parallel execution claims and points out the underlying communication complexity that diagrams omit.7:03–14:02 · The hosts pushing back 2/10 Performance Trade-Offs, Token Costs, and Evaluation Frameworks Dylan explains Anthropic's 15x token trade-off and five evaluation dimensions. Swyx demonstrates deep domain knowledge by explaining LLM judge consolidation mechanics, context engineering, and the safety contrast between read-only research and write-enabled coding agents.14:04–19:52 · The hosts pushing back 4/10 Why Cognition Favors Single-Agent Sequential Architectures for Coding Dylan demonstrates Cognition's sequential single-agent approach using a Flappy Bird example. Swyx questions whether the scenario is contrived, catches Dylan off guard by verifying the exact text in Cognition's post, and contextualizes Cognition's view against OpenAI research.19:53–24:15 · The hosts pushing back 4/10 A Decision Framework for Selecting Agent Architectures Dylan shares decision heuristics and Claude artifacts for agent selection. Swyx offers an authoritative counter-perspective on token economics, arguing that optimizing for cost is a trap because the price of intelligence drops 100x annually.

speaking balance: gold is the hosts, purple is the guest (3 minute bins)

0:00 · the hosts 29.2% · guest 70.8%0:00 · the hosts 29.2% · guest 70.8%3:00 · the hosts 2.7% · guest 97.3%3:00 · the hosts 2.7% · guest 97.3%6:00 · the hosts 13.8% · guest 86.2%6:00 · the hosts 13.8% · guest 86.2%9:00 · the hosts 24% · guest 76%9:00 · the hosts 24% · guest 76%12:00 · the hosts 40.8% · guest 59.2%12:00 · the hosts 40.8% · guest 59.2%15:00 · the hosts 15.8% · guest 84.2%15:00 · the hosts 15.8% · guest 84.2%18:00 · the hosts 41.6% · guest 58.4%18:00 · the hosts 41.6% · guest 58.4%21:00 · the hosts 31.7% · guest 68.3%21:00 · the hosts 31.7% · guest 68.3%24:00 · the hosts 67.4% · guest 32.6%24:00 · the hosts 67.4% · guest 32.6%27:00 · the hosts 0% · guest 0%27:00 · the hosts 0% · guest 0%
Sharpest disagreement ▶ 5:06 Dylan's skepticism on Anthropic documentation

Dylan playfully questions whether Anthropic actually executes sub-agent requests in parallel as stated, prompting Swyx to defend the company's technical integrity.

 Hardest push from the hosts ▶ 16:16 Swyx calls out contrived example and verifies source

Swyx directly challenges whether the Flappy Bird failure scenario was contrived and checks the blog text in real time to show Cognition literally cited it.

 Biggest teaching moment ▶ 9:38 Dylan breaks down Anthropic's eval dimensions

Dylan systematically breaks down Anthropic's five evaluation criteria and explains why marketing sites had to be downweighted against primary sources.

 The host holds their own ▶ 24:15 Swyx reframes token cost economics

Swyx challenges Dylan's framework criteria regarding token cost trade-offs, explaining that 100x annual intelligence deflation rewards builders who aggressively overspend on compute.
the scores for every segment, with the reasoning behind each
ChapterTopicThe hosts as informed peerGuest teachingGuest disagreementThe hosts pushing backWhy
Deconstructing Anthropic's Multi-Agent Architecture for Deep Research 5312 Dylan walks through Anthropic's deep research multi-agent architecture from his visual breakdown. Swyx lightly questions Dylan's skepticism regarding Anthropic's parallel execution claims and points out the underlying communication complexity that diagrams omit.
Performance Trade-Offs, Token Costs, and Evaluation Frameworks 7312 Dylan explains Anthropic's 15x token trade-off and five evaluation dimensions. Swyx demonstrates deep domain knowledge by explaining LLM judge consolidation mechanics, context engineering, and the safety contrast between read-only research and write-enabled coding agents.
Why Cognition Favors Single-Agent Sequential Architectures for Coding 6214 Dylan demonstrates Cognition's sequential single-agent approach using a Flappy Bird example. Swyx questions whether the scenario is contrived, catches Dylan off guard by verifying the exact text in Cognition's post, and contextualizes Cognition's view against OpenAI research.
A Decision Framework for Selecting Agent Architectures 7314 Dylan shares decision heuristics and Claude artifacts for agent selection. Swyx offers an authoritative counter-perspective on token economics, arguing that optimizing for cost is a trap because the price of intelligence drops 100x annually.

Statements from this episode (9)

Opinion
Davis: Claude Deep Research Outperforms OpenAI, Perplexity, and Gemini
“And time and time again, over the last couple of weeks, I found that Claude has by far outperformed the others. And I guess the definition of good for me right now is not just length, but also the number of sources and diversity of response.”
Dylan Davis Jul 5, 2025 ▶ 2:57
Assertion Partly supported
Anthropic finds multi-agent architecture outperforms single-agent baseline by 80%
“So the, in the blog post that Anthropik posted, they ran some tests and they noticed that the multi-agent structure outperforms the single agent structure by 80% based off a different variety of variables they measured.”
Dylan Davis Jul 5, 2025 ▶ 8:31
Assertion Supported
Davis: Multi-agent research consumes 15x baseline tokens versus 4x for single-agent
“So based off of a basic conversation, a single agent architecture for research is around four X, the number of tokens needed to achieve a research output. When you use multi-agent architectures, it's actually 15 X the number of tokens.”
Dylan Davis Jul 5, 2025 ▶ 8:58
Assertion Supported
Anthropic finds a single LLM judge outperforms five specialized judges
“They initially started with five LLM as judges. So each one of these points had their own LLM as a judge. They tested the ability and accuracy of that LLM as judge collective to judge, and it actually didn't perform as well as one. So they replaced all of thos…”
Dylan Davis Jul 5, 2025 ▶ 11:03
Insight
Davis: Multi-Agent Systems Struggle in Coding Due to Strong Sub-Task Coupling
“When using a multi-agent architecture versus a single agent for coding. And the main issue is that these are dependent upon each other. They're strongly coupled in the sense that when I take an action, that action then is impacts all the other sub-agents if it…”
Dylan Davis Jul 5, 2025 ▶ 15:40
Insight
Davis: Sequential Single Agents Succeed in Coding via Cumulative Context Passing
“By doing this approach with single agents, you're likely going to achieve More success, because you have all the context from previous agents. The reason it's beneficial is that all the actions taken previously are baked into that next agent's context, so it d…”
Dylan Davis Jul 5, 2025 ▶ 18:19
Insight
Davis: Multi-agent suits independent tasks, single-agent suits dependent pipelines
“So the first is, can you break the task into independent parts where they're not relying upon each other? So like, that's one abstraction away. Another one is, do you benefit from the chaos of having a multiple perspectives or different takes on a task that's …”
Dylan Davis Jul 5, 2025 ▶ 20:09
Assertion Partly supported
Swix: AI Inference Costs for Fixed Intelligence Fall 100x Annually
“The cost of intelligence for a given set of intelligence, let's say GPT-IV, let's say O-one, whatever, it is literally falling a hundred X over the course of one year.”
Shawn Wang Jul 5, 2025 ▶ 24:27
Insight
Swix: Cost-Unconscious Builders Win in AI by Building Ahead of Cost Curves
“So you should actually build your products ahead of where costs are so that by the time they are popular, actually your frontier. So it's actually, I think the efficiency thing trips up a lot of people in terms of Being cost and cost conscious and wasting a lo…”
Shawn Wang Jul 5, 2025 ▶ 24:41
Made with StarZero

Turn any episode into a week of clips.

This entire site, over 200 episodes transcribed, diarized, checked and made playable, runs on the StarZero media pipeline. Drop in your own episode and the podcast clipper finds the moments worth sharing, cuts them, captions them, and reframes them for every feed.

Try the podcast clipper Learn more