---
title: "Barry Zhang's Checklist: When to Build an AI Agent vs a Workflow"
date: "2026-11-05"
video_id: "7704764819936693517"
tiktok_url: "https://www.tiktok.com/@agentic.workers/video/7704764819936693517"
creator: "@agentic.workers"
category: "AI"
tags: ["ai agents", "workflows", "anthropic", "barry zhang", "building effective agents"]
excerpt: "Anthropic's Barry Zhang says build a workflow when the decision tree is clear. His checklist asks four questions: complexity, value, feasibility, error cost."
---
Barry Zhang of Anthropic argues that you should not build an AI agent for every task. If you can map the whole decision tree easily, you should build it as an explicit workflow, because that costs less and gives you more control. A TikTok re-post from @agentic.workers quotes this accurately, but the slide behind him lists four questions, and the clip covers only the first.

## Where the advice comes from

The clip is an excerpt from Zhang's talk "How We Build Effective Agents" at AI Engineer Summit 2025. The ai.engineer talk page credits him as the speaker, from Anthropic, and summarizes the talk as arguing that agents suit complex, ambiguous, valuable tasks, while predefined workflows give better cost control for clear decision trees.

During the talk he refers to Anthropic's blog post "Building effective agents" ("If you haven't read the blog post"). That post was published on December 19, 2024.

## Agent vs workflow: what's the difference?

Anthropic's post defines the two terms:

- **Workflows** are systems where LLMs and tools follow "predefined code paths."
- **Agents** are systems where the LLM directs its own process and tool use.

The post says agentic systems often trade latency and cost for better task performance. That tradeoff is the reason Zhang gives for not defaulting to agents.

## When should you build a workflow instead of an agent?

In the clip, Zhang says "The first thing to consider is the complexity of your task." He continues: "Agents really thrive in ambiguous problem spaces." Then: "if you can map out the entire decision tree pretty easily, just build that explicitly and then optimize every node of that decision tree."

His reason: "It's a lot more cost-effective and it's gonna give you a lot more control."

The TikTok description summarizes this as "Barry Zhang's practical test: if the decision tree is clear, build it explicitly." That is a faithful summary of what he says in the clip.

## The full "Should I build an agent" checklist

The slide on screen, titled Checklist: "Should I build an agent," has four rows:

1. **"Is the task complex enough?"** No leads to workflows. Yes leads to agents.
2. **"Is the task valuable enough?"** Under $0.1 leads to workflows. Over $1 leads to agents.
3. **"Are all parts of the task doable?"** No means reduce scope. Yes leads to agents.
4. **"What is the cost of error/error discovery?"** High means read-only or human-in-the-loop. Low leads to agents.

Zhang says "Here's our checklist," and the clip then covers only the first row before it ends. Someone watching only this excerpt could conclude that a clear decision tree is the only reason to choose a workflow. On the slide, task value, feasibility and the cost of errors are separate gates. Each one can push you toward a workflow, a smaller scope or a human check, even when the task is complex.

The dollar figures on the slide appear only as text. The clip does not explain them, and I could not read a full transcript of the talk, so it is unclear whether they are per-task budgets.

## When was the talk?

The description calls it "his April 2025 talk." April 2025 is when the video went public. Podwise lists the YouTube upload, "How We Build Effective Agents: Barry Zhang, Anthropic," on the AI Engineer channel with a publish date of April 4, 2025.

The live event was earlier. The ai.engineer page names it AI Engineer Summit 2025, and the official summit page gives the theme "Agents at Work." The organizer's Tito registration page describes the 2025 summit as "Agents at Work" at TimesCenter in New York, with the engineering sessions on Feb 21. The talk itself has no individual date on those pages, so February 2025 is the best-supported reading for when it was given on stage.

## Who posted the clip

@agentic.workers is a secondary account that re-posts conference excerpts with its own captions and badge. The description has no link and the clip has no product or call to action. The speaker's name is never said aloud in the clip. The attribution to Zhang comes from the description and matches the ai.engineer talk page.

## Bottom line

Zhang's advice in the clip is accurate and matches Anthropic's own writing: if a task's decision tree is easy to map, build an explicit workflow for cost and control. His slide adds three more questions about task value, feasibility and the cost of errors, which the clip skips. The talk was given at AI Engineer Summit 2025 in February and posted to YouTube on April 4, 2025.

## Sources

- [Building effective agents, Anthropic, December 19, 2024](https://www.anthropic.com/engineering/building-effective-agents): workflow and agent definitions and the latency and cost tradeoff.
- [How We Build Effective Agents, ai.engineer talk page](https://ai.engineer/talks/effective-ai-agents): speaker, event name and talk summary.
- [How We Build Effective Agents: Barry Zhang, Anthropic, Podwise, April 4, 2025](https://podwise.ai/episodes/3558886): YouTube publish date and title.
- [AI Engineer Summit 2025, Tito registration page](https://ti.to/software-3/aies-2025/): the "Agents at Work" TimesCenter schedule with Feb 21 engineering sessions.
- [AI Engineer Summit 2025, ai.engineer](https://www.ai.engineer/summit/2025): the summit's "Agents at Work" theme.


## Video description and verification notes

**The file is a 47.85 s vertical clip (1080x1920, MP4 container, 30 fps) with an HEVC Main video stream at 239,249 bit/s (1,435 frames) and an HE-AACv2 stereo audio stream at 44.1 kHz and 48,119 bit/s. ffprobe reports a container bit rate of 294,136 bit/s and a filesize of 1,759,415 bytes. metadata.json lists the format as bytevc1_1080p_294130-1 and leaves fps null, so the 30 fps figure comes from ffprobe. The post timestamp is epoch 1793905360, which is 2026-11-05 19:02:40 UTC and matches upload_date 20261105. The uploader is @agentic.workers (channel nickname "Agentic Workers", id 7493657629840524334). The audio credit is "original sound - Agentic Workers". The description is 176 characters: "Not every task needs an agent. Barry Zhang's practical test: if the decision tree is clear, build it explicitly. From his April 2025 talk on effective agents. #AIAgents #Workflows". At capture (2026-11-06 21:01:25 UTC) the video had 977 views, 21 likes, 0 comments, 2 reposts and 28 saves. I read six frames: 001, 005, 010, 015, 020 and 024. With extraction at 0.5 fps these sit at roughly 0, 8, 18, 28, 38 and 46 s. Frames 001 and 005 show a split layout with a close crop of a man in a tan overshirt and black t-shirt on a red circular rug at the top, a wide stage shot with the slide screen below, and a dark "Agentic Workers" badge at top right. Frame 001 shows a slide reading "Don't build agents for everything" under the burned-in caption "Don't build agents for". Frame 005 shows a lavender slide titled "Why not?" under the caption "valuable tasks. They shouldn't". Frame 010 shows a close-up of the speaker beside large white "AIE" letters, with the caption "like them and they're a great". Frames 015, 020 and 024 show a slide titled Checklist: "Should I build an agent" with an Anthropic logo, laid out as a four-row table. The caption reads "The first thing to consider is" in frame 015, "easily, just build that" in frame 020 and "control." in frame 024. Frame 024 also shows an "AGENT ENGINEERING https://ai.engineer" banner and an "AIE" logo. No terminal, code, demo or product UI appears in any frame I read. The transcript has 18 segments and 125 words. The speaker says agents suit complex, valuable tasks and that workflows are concrete and available today. He says that if the whole decision tree is easy to map, you should build it explicitly and optimize each node, because that is more cost-effective and gives more control. The last segment ends at 46.82 s, inside the 47.85 s audio, so the transcript is not hallucinated past the end.**

## The spoken claim: map the decision tree, build it explicitly (verified, with one row of four)

At 27.34 s the transcript says "The first thing to consider is the complexity of your task." At 31.72 s: "Agents really thrive in ambiguous problem spaces." At 35.08 s the speaker continues "if you can map out the entire decision tree pretty easily, just build that explicitly and then optimize every node of that decision tree." At 43.34 s he gives the reason: "It's a lot more cost-effective and it's gonna give you a lot more control."

This matches the two Anthropic-side sources I could read. The [ai.engineer talk page](https://ai.engineer/talks/effective-ai-agents) summarizes the talk as arguing that agents suit complex, ambiguous, valuable tasks, while predefined workflows give better cost control for clear decision trees. Anthropic's [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) post, which the speaker refers to at 14.94 s ("If you haven't read the blog post"), defines workflows as systems where LLMs and tools follow "predefined code paths" and agents as systems where the LLM directs its own process. It says agentic systems often trade latency and cost for better task performance. The video's advice and its cost-and-control rationale agree with both.

The TikTok description calls this a "practical test". That framing fits one row of a four-row checklist, not the whole checklist, as the next section shows.

## What the video skips: the slide has four questions, the narration covers one

The frames from 015 onward show the slide in full. The four rows, quoted from the burned-in slide, are "Is the task complex enough?" (No leads to Workflows, Yes to Agents), "Is the task valuable enough?" (under $0.1 leads to Workflows, over $1 to Agents), "Are all parts of the task doable?" (No means reduce scope, Yes means Agents) and "What is the cost of error/error discovery?" (High means read-only or human-in-the-loop, Low means Agents).

At 26.32 s the speaker says "Here's our checklist," and the audio then covers only the first row, complexity. The clip ends at 46.82 s with the first item still the topic. A viewer who sees only this clip, or only the description, could conclude that a clear decision tree is the sole reason to choose a workflow. On the slide, task value, feasibility and error cost are separate gates, and the description does not say so. This is a framing gap in a short excerpt, not a false statement. The description's own wording ("if the decision tree is clear, build it explicitly") is faithful to what the speaker said.

The dollar thresholds on the slide ("<$0.1" and ">$1") appear only visually. I could not read a full transcript of the talk, so I cannot say how the speaker explained them or whether they are per-task budgets. I treat the figures as slide text that the clip does not discuss.

## Event and date: "April 2025 talk" is the upload month, not the stage date (partial correction)

The description says "From his April 2025 talk on effective agents." Podwise lists the YouTube upload, titled "How We Build Effective Agents: Barry Zhang, Anthropic", under the AI Engineer channel with a publish date of [April 4, 2025](https://podwise.ai/episodes/3558886). The talk page on ai.engineer names the event "AI Engineer Summit 2025" and gives no talk date.

Frame 010 shows an "AIE" logo, and the frame 001 wide shot shows a podium sign that appears to read "Agents at Work". The organizer's [Tito registration page](https://ti.to/software-3/aies-2025/) describes the 2025 summit as "Agents at Work" at TimesCenter in New York, with the engineering sessions on Feb 21. The [official summit page](https://www.ai.engineer/summit/2025) states the theme "Agents at Work." Sources I found put the in-person summit in February 2025, not April. I could not open a page that gives the exact talk date or the exact upload date from first-party YouTube metadata, so I treat "stage date in February 2025, public video April 4, 2025" as the best-supported reading. The stage date is inferred from the event schedule, not from a talk-level record.

The blog post the speaker mentions predates both: the fetched Anthropic page shows a publication date of December 19, 2024. The clip's narration ("If you haven't read the blog post") therefore points to a post that was already about four months old when the video was uploaded.

## What the speaker and uploader are, and what the clip sells

The speaker's name is not spoken in the transcript. The attribution to Barry Zhang comes from the description and is corroborated by the ai.engineer talk page, which credits him as the speaker, affiliated with Anthropic. Frame 001 also shows an Anthropic shirt logo and the slide footer reads Anthropic. The Anthropic blog post lists him alongside a co-author in its acknowledgements.

@agentic.workers is a secondary account re-posting a conference excerpt with its own badge and captions. The description carries two hashtags and no link, and the audio credit is the account's own "original sound". No call to action, product or guide appears in the transcript or in the six frames I read, so there was no outside resource to follow beyond the talk and the blog post. I did not fetch any tiktok.com page.

## Key Takeaways

- **Verified:** At 31.72 to 46.82 s the speaker says to build an explicit workflow when the decision tree is easy to map, citing cost-effectiveness and control. The ai.engineer talk summary and Anthropic's blog post support that guidance.
- **Partial correction:** The description calls the checklist a test based on a clear decision tree. The slide in frames 015 to 024 has four questions (complexity, value, feasibility, error cost), and the narration covers only the first.
- **Partial correction:** The description says "April 2025 talk." The sources I found tie the in-person event to February 2025 (AI Engineer Summit, "Agents at Work"), and the YouTube upload is listed as April 4, 2025.
- **Clarification:** The slide thresholds of under $0.1 and over $1 are visible in the frames but not spoken in the clip, and I could not read the full talk to see how they were explained.
- **Context:** The clip is a re-post by @agentic.workers with 977 views and 21 likes at capture. It contains no demo or terminal, and the transcript ends at 46.82 s inside a 47.85 s audio track.

## Resources

- [Building effective agents, Anthropic, December 19, 2024](https://www.anthropic.com/engineering/building-effective-agents) establishes the workflow versus agent definitions, the latency and cost tradeoff, and the blog post the speaker mentions.
- [How We Build Effective Agents, ai.engineer talk page](https://ai.engineer/talks/effective-ai-agents) establishes the speaker, the event name "AI Engineer Summit 2025" and the talk's summary of complexity and cost control.
- [How We Build Effective Agents: Barry Zhang, Anthropic, Podwise, April 4, 2025](https://podwise.ai/episodes/3558886) establishes the YouTube publish date and title.
- [AI Engineer Summit 2025, Tito registration page](https://ti.to/software-3/aies-2025/) establishes the "Agents at Work" TimesCenter schedule, including the Feb 21 engineering sessions.
- [AI Engineer Summit 2025, ai.engineer](https://www.ai.engineer/summit/2025) establishes the summit's "Agents at Work" theme.
