Few recos for the @OpenAI agents team:
1. Support editing the agent within a chat session.
2. Bug: Chat sessions repeat previous asks.
3. Group agent chats in the sidebar.
4. Bug: New chat shortcut spawns chat without agent context.
5. Improve agent schedule flow
I kind of need @github integrated in everything we vibe generate - design, prototypes, presentations, roadmaps, ...so painful to not have it universally available across the agent harnesses.
@bcherny@pmarca@bcherny we're running a similar experiment at https://t.co/t8DAJEVgmQ. We're in early stages so we tend to be more in 1/2/3. We've also collapsed these into a single person and we're finding a lot of success with it.
Impressed with @grok 's video generation capabilities. One prompt including instructions on Hindi content and it created exactly what I was looking for. Well done!
Claude Desktop won't tell you how much of your context window your MCP servers eat before you even type "hi." Built a CLI that shows you. Also see what changed since the last time you took a snapshot. Open source, MIT. https://t.co/n8z67V8HMm
@claudeai
@AndrewYNg the hybrid role of an engineer with a product sense is what I believe will be the most successful in moving small to medium capabilities. Surround that person with a a strategy context and an infrastructure that lets them experiment and move fast. Combine this with an agent driven process and it starts to integrate with PMM, compliance, security, legal, etc.
@karpathy@karpathy how do you think about the state machine of the wiki - something in draft mode (so it doesn't change the permanent structure) or if multiple folks are contributing then review/approve.
What would be cool is if @Microsoft@Copilot provided a prebuilt harness that:
1. Leverages Teams channels/group chats for authoring and distribution permissions
2. Provisions and deploys agents and artifacts to the IT approved/managed cloud
3. Plugs into the PC and LOB device policies to executes on device tasks
4. Embeds group comms so multi authoring is viable
5. Provides access to M365 artifacts and tools
6. Auto versions my agents/artifacts through GitHub
...and make the entire experience for a non engineer consumer like.
That would be an interesting Copilot harness!
@chamath I like where the factory is headed in terms of having an agent manage all things related to building product.
Having led products at Microsoft, SoFi, and Carta, I know first hand that the process of building a product is messy. I agree that the artifacts are spread all over the place.
However, in my view you should separate the 8090 agent from wiki management. How the PRDs and blueprints are managed for the agents to reason over is the LLMs problem. From time to time as a user I can ask the LLM to show its internal representation or I can directly take a raw input source and modify it but outside of that I believe the only interaction should be through the 8090 agent. As meetings or chats happen where decisions are made, the agent should pull me in and confirm how that changes things and then update its internal representation.
From there it can then influence the work items and/or code.
This enables more purpose specific tools to be leveraged and as long as the 8090 agent is aware of those tools you're good.