Chase

@jess__yan Thanks for sharing. My work has evolved similarly. One of the agents we’ve built is a branding agent that helps get the prototypes I build to high fidelity in alignment with design and brand standards

I recently had a chance to reflect on how the product role has evolved along the AI exponential. Here's a look into my experience building Claude Managed Agents, plus practical examples of the DIY'ed agents currently unlocking my productivity. claude.com/blog/product-d…

@jess__yan My team recently built a similar service internally at Intuit. Looking forward to diving into your product!

Turns out the best way to handle lots of demand for Claude Managed Agents is actually... spinning up a few Claude Managed Agents. Keep the feedback coming!! 🌊

@atr0t0s @levelsio This is the right take

MCP is very much alive. Growing even. Really valuable for discovery, auth, bidirectional streaming, and portability across LLM providers without custom integrations. MCP provides a standardized AI-friendly abstraction layer on top of existing APIs. For an agent to hit an API it needs to learn that API specifically - if you need to change API providers for the same function, their API will be different. With MCP more than one API provider that offer similar or overlapping functionality can be hidden behind the same standardized MCP interface, so the LLM only needs to learn one consistent set of tool names, descriptions, calling patterns, etc As usual the truth is more nuanced than what a tech influencer will tell you in their customary ragebait posts.

Thank god MCP is dead Just as useless of an idea as LLMs.txt was It's all dumb abstractions that AI doesn't need because AI's are as smart as humans so they can just use what was already there which is APIs

@JaySahnan Or is engineering a sales problem?

GTM is no longer a sales problem its an engineering one

@stephenhaney Loving this. Using right now and enjoying the experience. Much better than alternatives I’ve tried

Hello! Today we're releasing Paper Desktop Paper is now a canvas for Cursor, Claude Code, Codex. Any agent can read and write html to Paper. • push or pull from your codebase • pull real data from anywhere • less work, more design What will you ship? Sound on 🎶

@codyschneider Why would a pm spend time prototyping something users wouldn’t want?

talked to my friend, he's a PM at a big company it told me that almost everybody in the product side of the work is using AI for prototyping but they're just building things things that nobody wants it's engineering again building for the sake of build not because the buyer wants it if the cost of product goes to zero if the cost of code goes to zero building the thing is what doesn't constrain you anymore building the right thing is what constrains you and the right thing is what the market wants to buy and there are so few people that know how to identify what the market wants to buy

@realmadhuguru Excellent take. Resonates with what I’ve found to work well at Intuit

The best enterprise AI implementations I see: Pair workflow experts + people with great product sense (rare). Understand workflows deeply - tools, manual steps, coordination across systems, hallway conversations. Common gaps to address: - Codify institutional memory (in people's heads) - Codify undocumented processes - Clean existing docs (which of the 3 HR policy docs on the intranet is the latest?) Build: spec workflows, representative evals, systems that auto-improve as models evolve. Critical miss: treating AI like traditional software. Models evolve every 6-8 weeks. Need continuous iteration, not annual roadmaps.

@Saboo_Shubham_ Good read, thanks for writing. I do something pretty similar, I use three agents in cursor to spin up different versions of a prototype. Love their multi-agent feature, so fast and super helpful in identifying edge cases/fleshing out solutions

@langchain Do people actually find success with thousands of lines of agent prompts?

@rahulgs @rickasaurus Agreed, fuzzy search is fine in non-crit spots, no real reason to fixate guardrails. I've experienced embeddings hallucinating connections in specific domains so I try to have my teams be more vigilant. I find top-k includes drifts / misaligned “relevance” with semantic search

Fuzzy search is typically used to rank suggestions in non-critical places. In our experience, a lot of these cases are improved by using semantic search. Guardrails aren't really needed because you wouldn't or shouldn't have been using fuzzy search in a place that needed it anyway. Thoughts?

yes things are changing fast, but also I see companies (even faang) way behind the frontier for no reason. you are guaranteed to lose if you fall behind. the no unforced-errors ai leader playbook: For your team: - use coding agents. give all engineers their pick of harnesses, models, background agents: Claude code, Cursor, Devin, with closed/open models. Hearing Meta engineers are forced to use Llama 4. Opus 4.5 is the baseline now. - give your agents tools to ALL dev tooling: Linear, GitHub, Datadog, Sentry, any Internal tooling. If agents are being held back because of lack of context that’s your fault. - invest in your codebase specific agent docs. stop saying “doesn’t do X well”. If that’s an issue, try better prompting, agents.md, linting, and code rules. Tell it how you want things. Every manual edit you make is an opportunity for agent.md improvement - invest in robust background agent infra - get a full development stack working on VM/sandboxes. yes it’s hard to set up but it will be worth it, your engineers can run multiple in parallel. Code review will be the bottleneck soon. - figure out security issues. stop being risk averse and do what is needed to unblock access to tools. in your product: - always use the latest generation models in your features (move things off of last gen models asap, unless robust evals indicate otherwise). Requires changes every 1-2 weeks - eg: GitHub copilot mobile still offers code review with gpt 4.1 and Sonnet 3.5 @jaredpalmer. You are leaving money on the table by being on Sonnet 4, or gpt 4o - Use embedding semantic search instead of fuzzy search. Any general embedding model will do better than Levenshtein / fuzzy heuristics. - leave no form unfilled. use structured outputs and whatever context you have on the user to do a best-effort pre-fill - allow unstructured inputs on all product surfaces - must accept freeform text and documents. Forms are dead. - custom finetuning is dead. Stop wasting time on it. Frontier is moving too fast to invest 8 weeks into finetuning. Costs are dropping too quickly for price to matter. Better prompting will take you very far and this will only become more true as instruction following improves - build evals to make quick model-upgrade decisions. they don’t need to be perfect but at least need to allow you to compare models relative to each other. most decisions become clear on a Pareto cost vs benchmark perf plot - encourage all engineers to build with ai: build primitives to call models from all code bases / models: structured output, semantic similarity endpoints, sandbox code execution. etc What else am I missing?

@rahulgs @rickasaurus I’m a proponent of semantic search via embeddings too but question whether topk is the right metric to look at without considering guardrails

@rickasaurus in any mature codebase there are likely instances of sorting/suggestions based on text search. we replaced a bunch of those cases with semantic search and saw immediate increase in topk retrieval rates

@ankurnagpal How does rebalancing work with this strategy?

2025 was the first year I switched to direct indexing from investing in index funds Results: - Up ~20% for the year (consistent with index funds) - $135K in usable tax losses Since I had a large capital gain, this will save me at least ~$66K in taxes this year alone

11/17 The only framework that works consistently: 1. What problem are we solving? 2. For who specifically? 3. How will we know it worked? 4. What's the simplest version? 5. What could go wrong?

12/16 The best "roadmaps" I've seen: 3 months: Specific commitments 6 months: General direction 12 months: Strategic themes The further out, the vaguer it gets. That's honest planning.

@garrytan I’m doing it in most on my work. I use em dashes regularly so people sometimes think my writing is AI. A few typos goes a long way

How long before people intentionally put in typos to make sure people believe an LLM didn’t write a thing?

Easily the best breakdown of use case I've seen

@reidhoffman Think it’s a mixture of bloodbath and enablement. AI is currently great at reducing manual workflows but it’s tough to exceed that. Manual work should be reduced. Creative, strategic work will hopefully be the forefront of human labor

Some AI industry leaders are predicting white-collar bloodbaths. Even the most inspirational advice to new graduates lands like a Band-Aid on a bullet wound. Some thoughts on new grads, and finding a job in the AI wave:

@tomkrcha Would like to give this a go

Introducing Design Mode for Cursor! 🧑‍🎨🤖 Vibe code and vibe design in one place, stay in the flow. Uses the same existing AI agent that's in your favorite coding tool, same context, intertwined.

@chrism_biz @NickADobos They do

@NickADobos I find it annoying, mostly. Now, when I ask a simple question, it gives me unrelated coding answers based on things that are from months ago. They should give you the ability to save individual chats to memory, not just blanket it all in automatically.

ChatGPT memory is definitely making me realize having a big history of conversations is a huge huge huge asset. All the other apps with way less usage have an uphill battle to build that personalized history

@jayabdulraman @nurijanian Only technical PMs. Depends on your relationship with developers

@nurijanian Hope you’re faring well. Here’s a question from me, a new bee product manager. Do product managers engineer a solution like figure out the how to present to engineers?

The toughest part of moving to product? Not strategy, stakeholders, or politics. It's staring at a blank PRD, unsure if it's right. I faced this too, until I found a system. Here's how to overcome documentation paralysis 🧰👇