Mo

@meb

587 Following

591 Followers

Mo pfp

How I would define an AI agent at the start of 2025: In abstract, a piece of software that can achieve a certain goal, and uses AI to navigate the fuzzy path of achieving that goal. More specifically, a combination of - Goal - Context management - Underlying model - Tools (datastores and functions with effects) - Prompting & self-prompting strategies (including delegating to other agents) Example: A geocoding agent Goal: Parse content from a sentence and geocode it Context Management: Trimmed sliding window. We only need the very most recent messages to get the job done. Underlying model: A simple model like 4o mini is more than enough Tools: A function to search cities with inbuilt query match ranking, and a function with access to a Google Maps geocoding API Prompting: Single turn. The agent should be able to get the job done in a single exchange, or ask the user to clarify the location if needed

0 reply

0 recast

4 reactions

Mo pfp

Until recently, gpt 4o mini has been my go to for lots of agents given its low cost and latency. Did some tests today with Gemini 2.0 flash, which somehow is even less expensive, but supposed to be full 4o parity and got amazing results, especially with it following instructions very well. Shoutout to @alexpaden for sharing the tip on xml-ish tags for your prompts. Seems to be more machine readable, but also makes it way easier to structure

1 reply

0 recast

1 reaction

Mo pfp

Today, for the first time in a while, I solved an issue by going from stackoverflow, to Reddit, to the official docs. Nostalgia vibes

1 reply

0 recast

1 reaction

Mo pfp

First time vibe coding yesterday (swift in cursor). Very cool experience and moved ahead incredibly fast compared to the old days of reading the docs, and moving up the stack from "hello world" when working with new languages. I don't think all devs are cooked yet, but this has already passed the level of the average "copy paste stack overflow without caring" dev. What I'm enjoying: 1. Being able to spend more time in product land, versus figuring out syntax. 2. Reducing the barrier to porting my engineering skills to a different language. I think LLMs will be an awesome supporter for the mobility of experienced engineering talent between language ecosystems. 3. Setting up boilerplate and screens way faster, so I can focus on iterations Speedbumps: 1. Architecture. Getting the AI to follow rules such as decomposing screens into different files, and explain certain design decisions it makes. 2. Verbosity. Feels like too much written some times. I generally have to prompt for simplicity.

0 reply

1 recast

1 reaction

Mo pfp

Looking to get into iOS app development Go for swift or react native? Not intending to publish to android (for now)

7 replies

1 recast

4 reactions

Mo pfp

Saved some time using AI to translate an English doc to French. The canonical google French version for Farcaster is "Farcasteur"

2 replies

0 recast

5 reactions

Mo pfp

Feature request inspired by situation I just had; Can we get LinkedIn post embeds like we have with Twitter? Pros: - Enable people targeting clients outside farcaster world to reshare content here - Extend the farcaster social graph back into places like LinkedIn - Build the bridge between normies and crypto natives

2 replies

0 recast

2 reactions

Mo pfp

AI Knowledge Graph productivity hack Send me some <3 on LinkedIn https://www.linkedin.com/feed/update/urn:li:activity:7311890866313297920/

0 reply

0 recast

1 reaction

Mo pfp

Following on yesterdays chat about Knowledge Graphs I discovered Microsoft recently released a framework that reverse engineers knowledge graphs from unstructured text https://youtu.be/6vG_amAshTk?si=iG31IpbvsCZZ9dx4

1 reply

0 recast

11 reactions

Mo pfp

This week I’m going to hit my running goals higher ↑

0 reply

0 recast

1 reaction

Mo pfp

What are some good example projects where an agent is leveraging MCP client? Not looking for claude desktop, but an actual AI agent that has a layer that fetches tools from MCP before issuing a chat completion that may return tool_call

1 reply

0 recast

5 reactions

Mo pfp

See my reaction times in Reflex Rush on the @atlas frame: 522ms 392ms 403ms 387ms 395ms Average time: 420ms

1 reply

0 recast

0 reaction

Mo pfp

Unwinding after a long day. Relaxing by reading the drama between Rippling and Deel. Wishing myself and all the founders here to reach a level where we need to honeypot an employee that we suspect of corporate espionage https://x.com/parkerconrad/status/1901615179718406276?s=46&t=heYf7m0kaq62UF_MFK0e6g

1 reply

0 recast

4 reactions

Mo pfp

Writing onchain RBAC enforcement contracts is my comfort zone

1 reply

3 recasts

2 reactions

Mo pfp

0 reply

0 recast

1 reaction

Mo pfp

0 reply

0 recast

3 reactions

Mo pfp

People criticise wrappers, but I just discovered that Bankr bot is itself a wrapper around Cowswap, which itself wraps a number of trading protocols. There is something to be said for convenience and packaging something people already do into an easier way

2 replies

0 recast

2 reactions

Mo pfp

Cursor .mdc files for supercharging your workflows Think of these as super atomic ways to guide cursor on how to build. There are 2 important fields - Description: Cursor evaluates this description based on your request, and loads it if necessary - Auto-attach: Instructions always loaded when file path matches I recommend adding descriptions for your own sanity, but making stringent use of auto-attach in a proper granular way. My current use case; - Consistent styling (function components instead of const X = () => {}) in React - Templated setup for agents: Spinning up new agents following existing patterns

1 reply

1 recast

3 reactions

Mo pfp

I still don't fully understand MCP, but I do understand ERC20s

0 reply

0 recast

2 reactions

Mo pfp

Following on yesterdays, MCP discussion, a very interesting X thread about this In brief; MCP makes sense as a clearing house for capabilities. Agents don’t need to rebuild the tool integrations, they can just access existing capabilities with standard auth. This reminds me of Amazons policy that teams and their products can only communicate over APIs https://x.com/abacaj/status/1897040132974538835?s=46&t=heYf7m0kaq62UF_MFK0e6g

1 reply

0 recast

3 reactions