Claude's /goal Feature Just Exposed A Real Challenge With AI Agents

Help me think through how this episode applies to my situation. Start by asking what I am trying to change. Separate the episode's claims from your suggestions, and say when the notes do not support a claim. Use the transcript to find passages, then check the audio before quoting. Transcript: https://yuvalyeret.com/podcast/episodes/claudes-goal-feature-just-exposed-a-real-challenge-with-ai-agents/transcript.md ## Published episode notes Claude and Codex have a relatively new feature , `/goal` , that lets you set a completion condition and keep the AI running autonomously until it's met. This capability exposes a real challenge with agentic AI and explains why its shifting the bottleneck to its assymetric capabilities in different stages and domains of value creation. 00:09 The new /goal capability , what it does and how it works 02:07 What the official examples reveal: all output, no outcome 03:02 The missing examples , what outcome-oriented goals would actually look like 04:00 Output vs. outcome: why the gap matters for AI impact 04:45 The real bottleneck , observability and closing the feedback loop 05:45 Live demo: setting an outcome-oriented /goal on a blog post 08:30 What the loop did , changes made, open items, what's next Are your AI goals output-oriented or outcome-oriented? What are you doing to enable outcome-oriented goals in your AI sessions? And as a final thought - What would happen if your human teams were empowered to seek a /goal? Insights on Scaling AI from Activity to Impact Join the Activity to Impact Conversation on Linkedin Goal-Based Loop Engineering: From Spec-Driven to Outcome-Driven – https://yuvalyeret.com/blog/ai-agent-completion-goals-aim-at-outcomes ## Transcript Automatic transcript of the published podcast audio. Recognition errors are possible. Speakers are unlabeled; do not attribute a passage to Yuval or a guest without checking the audio. Episode: https://yuvalyeret.com/scaling-ai-podcast/claudes-goal-feature-just-exposed-a-real-challenge-with-ai-agents/ Source RSS GUID: 955fd625-081c-4bb3-b50f-29ad508824a0 Source: published Riverside RSS audio enclosure Transcription: faster-whisper base.en, English ## Transcript [00:00:00] One way to help scale AI from activity to impact is to actually specify [00:00:07] the impact that you want to see. I just received an email from Claude that [00:00:15] includes a couple of new feature updates. One thing that caught my eye was [00:00:22] keep Claude working towards a goal, which is a new, relatively new capability, I guess, in Claude, [00:00:31] which basically means that you can guide Claude in running a route loop and automation loop by [00:00:42] setting a completion condition using a goal. So essentially, the way it works is you set a completion [00:00:50] condition with slash goal. And Claude, or CODEX, has the same thing, keeps working across turns [00:00:57] until the condition is met. One of the interesting things to think about is what would that completion [00:01:04] condition be? So if we think about the fact that Claude would keep working towards that goal without [00:01:11] prompting, after each turn, what it does is it checks. And here, it uses a small fast model to check [00:01:19] whether the condition holds is met. If not, it will start another turn instead of returning control to [00:01:27] you. That's a higher level of agency. It's sort of a mini route loop, a mini loop of running the [00:01:36] model continuously until you get to a certain point. The examples here are more of an example of [00:01:48] what can be done. It's more interesting to actually look at what don't we see here. [00:01:54] So for example, migrating a model to a new API until every call site compiles and tests pass, [00:02:02] rate implementing a design doc until all acceptance criteria hold, splitting a large file into focus [00:02:09] models as some sort of tech, that improvement refactoring, working through a labeled issue deck [00:02:16] slot until the queue is empty, which is kind of a meta loop. Because for each one of these issues, [00:02:25] they would probably need to be a goal in achieving that goal. If you look at this list, [00:02:34] one of the things that I see here is all of these are technical activities. It doesn't really show [00:02:42] here examples of a web page is working well. That feature is desirable. People use the feature. [00:02:53] People use the functionality. We created a presentation that people actually understand. [00:03:00] We created an image that is useful for us. We've done something that achieves the change that we want [00:03:09] to see for the people that would use it. And those are the sort of goals that we should be [00:03:17] aiming at. That's the intent of the work that we're doing with AI. All of the examples that we have [00:03:23] here, yeah, they're serving us, but there are examples of output. There are not examples of [00:03:30] outcome. Output doesn't create impact. It's only if that output is what we are doing here. [00:03:38] The tactical goals, if you want to call it that. Turn into outcome. Turn into something that [00:03:46] moves the needle, changes how people work, that it has a chance to achieve impact. I really like [00:03:54] this approach, but I'm curious to see how many people are going to use it for mechanical output [00:04:05] oriented goals. What will need to be true in order for us to use this for more outcome oriented [00:04:15] goals? First of all, we need to learn how to frame outcome oriented goals effectively, but the [00:04:22] other piece of this is how would this loop run? So how would this model be able to actually close [00:04:33] the loop on? Did we achieve the target condition? How would we observe? And that's where I think a [00:04:40] lot of the work will need to take place. How do we move, tackle the bottleneck, the new bottleneck, [00:04:49] which is validating that we've delivered value through new observability, new ways to close the [00:04:58] feedback loop. Only when we do that, will we be able to give a gentiKI real outcome oriented goals [00:05:08] and unleash agents on real problems, not just doing technical tasks along the way? Chris, [00:05:19] how you're using goal? Meanwhile, I'm starting to play with it in my world. So just as an example, [00:05:32] this is a blog post that I'm working on. I think it's in reasonable shape. It's about my AI bottleneck [00:05:38] wasn't tokens. It was finishing my AI threads board. But what I've done here is set a goal to [00:05:49] cloud that my blog post should be well aligned with my activity to impact bottleneck point of view [00:05:58] and an effective blog post that will show up in answer engines and work well as a link to in [00:06:06] sub stack newsletter. And then essentially, I see here that cloud started to work on this. [00:06:14] It's done some work. It made some changes to the flow of the language that I came up with. [00:06:22] It changed the closing line. It explains why it thinks it serves the goals, how it connects to the [00:06:31] bottleneck point of view. It's not surprising that it's already aligned. That's how I wrote this [00:06:37] article. It explains what has it done for answer engine optimization. It added some headers that are [00:06:47] question oriented that the first TLDR at the top, which provides a quick, [00:06:56] portable, liftable answer blog. And also thinks it's a newsletter fit. It does have an open item, [00:07:05] which is the image. Cloud cannot generate an image. I didn't teach it to open a subagent with code [00:07:14] X to generate images yet. And then let's see what has it done. It continued in the loop. [00:07:24] It continues in the loop. And what it's doing is committing this even without the cover image. [00:07:34] I guess to render something. We'll be curious to see how it works. So anyhow, I'll continue to work [00:07:42] and explore with using goal on my AI threads. And let me know what you're doing with it. And I'll [00:07:54] and are your goals, outcome oriented, output oriented, what are you doing in order to enable [00:08:00] outcome oriented goals in your AI sessions. This is killing AI from activity to impact. [00:08:07] But with a quick update on orienting AI harnesses around goals. See you soon. ## Source boundary These are the published show notes from the podcast feed. They are a starting point for discussion, not a verbatim record of the conversation. The transcript is machine-generated and may contain errors or unlabeled speakers. Check the audio before quoting anyone.