andyrewlee

agent orchestrator connoisseur

[andyrewlee @andyrewlee]

@dimagrossman my experience is from an engineer perspective integrating with other apps. most users used it as a generalist agent to manage other apps via mcps.

[andyrewlee @andyrewlee]

i’m a top 1% user of claude managed agents.

i have mixed feelings.

it’s the best designed managed agents platform.

but it can never be the best because it isn’t harness or model agnostic.

[andyrewlee @andyrewlee]

@ishaansehgal 👀

[andyrewlee @andyrewlee]

fable is a better model but i'm using 5.6 10x more
1. no threat of being removed
2. no 5 hour limits
3. no auto switching to weaker model
4. comparable output
5. better for goals
6. generous resets
7. active sessions continue after usage limit
8. codex cli is open source

[andyrewlee @andyrewlee]

current setup
- fable 5 high for active work (claude code cli)
- gpt 5.6 sol max for goals (codex cli)
- glm 5.2 for personal assistants (openclaw, hermes)

[andyrewlee @andyrewlee]

@Fred_ayyyy it says everyone 🤷

[andyrewlee @andyrewlee]

omg claude usage reset

[andyrewlee @andyrewlee]

there’s debate around monoagent or microgents future.

future probably looks close to this.

end user interacts with a monoagent but under the hood it’s an orchestrator of microagents.

[Sakana AI @SakanaAILabs]

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API.

Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls.

Try it: https://t.co/hhO6qTawgb 🐡

[andyrewlee @andyrewlee]

first time not missing frontier models as i use glm 5.2 with hermes and openclaw

to maximize return on tokens
- plan with fable (i miss you)
- execute with glm 5.2
- review with gpt 5.5

[ALEX PARK • 박지흔 @alexpark]

Presenting my newest AI short film:

DON'T BLAME THE COMPUTER.

Made entirely with AI - about why that's not the point.

[andyrewlee @andyrewlee]

here’s how I’m looping responsibly:

1. $autoreview keeps looping until all issues are addressed

2. discuss intent with agent so it can define its own /goal and loop to completion

3. set up scheduled automation to inspect production traces and fix issues

[andyrewlee @andyrewlee]

focus on helping agents fix bugs

https://t.co/kyEWHKlEfh

[Alvin Sng @alvinsng]

If agents can’t solve a bug, your observability is broken.

Close logging gaps. Give agents access to logs, databases, and the codebase. Create the right skill files.

Humans should fix the setup, not the bug

[andyrewlee @andyrewlee]

3 recent posts that made me a better developer

[andyrewlee @andyrewlee]

i never write my goals

https://t.co/ywOu6hVecX

[jason @jxnlco]

tips for codex goals

sure you can use /goal but it also has a set_goal() function

its almost better to prompt the model to set its own goal, it will likely write a better prompt than you

[andyrewlee @andyrewlee]

uber engineers will never get to use workflows

[Thariq @trq212]

Workflows are the biggest upgrade to Claude Code’s capabilities since skills and subagents.

I dove deep into it with @sidbid to figure out best practices, examples and more.  I’m particularly excited about the non-technical tasks it enables for Claude Code.

[andyrewlee @andyrewlee]

current workflow for projects

1. create rfc on chatgpt with 5.5 pro
2. create linear project with rfc
3. open codex cli 5.5 high
4. set goal to implement linear project as stacked diff
5. test at top of stack using computer use
6. run autoreview skill on each pr
7. merge stack

[andyrewlee @andyrewlee]

@pashmerepat surprised figma didn't trademark "code mode"

[andyrewlee @andyrewlee]

@kaiyuhsu which clients handle it the best? let’s say there’s 5 mcps connected with 500 tools total

[andyrewlee @andyrewlee]

i’m finally starting to see the light

mcp > cli

dynamic tool discovery is a must however when using multiple mcps