Andrew Sinagra

Building @coordHQ - the agent farm platform. CoFounder @nimlabs. Creative Director @NTROPIC

[Andrew Sinagra @sinagra]

@asha_shar I think @GamersNexus needs to do a case review on this one.

[Andrew Sinagra @sinagra]

I gotta say... the hand drawn aspect to Claude's icons and animations is a brilliant move to bring a human feel to agents and their actions.

[Andrew Sinagra @sinagra]

@thdxr [email protected]

[Andrew Sinagra @sinagra]

Opus 5.5 feels like I get to work with an old friend again instead of someone who is trying to use big words to prove how smart they are all the time. LOL

[Andrew Sinagra @sinagra]

@VadimStrizheus Fable is for orchestration. Use Opus for implementation. I can run multi-day sessions from a Fable orchestrator and reviewer without getting close to the weekly limit on the Max (20x) plan.

[Andrew Sinagra @sinagra]

@SebastianRoehl Whenever I’m waiting on agents I switch over to designing the next surface or refining existing ones in Figma. It’s a good mental break rather than spinning too many plates at once.

[Andrew Sinagra @sinagra]

@thomasgauvin You don’t need a harness. You need an orchestrator.

[Andrew Sinagra @sinagra]

So both Claude and Codex are down... my spidey are senses tingling.

[Andrew Sinagra @sinagra]

I get it... we're holding

[Andrew Sinagra @sinagra]

@amorriscode Looking forward to it!

[Andrew Sinagra @sinagra]

This is essentially my orchestration arc. 3 skills all working together.

/orchestration-arc
/arc-issue
/adversarial-gauntlet

The orchestration arc does the brainstorming, builds the plan, determines the PRs and then writes each to a GitHub issue. Then it kicks off each GitHub issue as an independent session via /arc-issue which explains to the session it is running in an arc, how to pick up the GitHub issue, and uses inter-session messaging to communicate all questions and final status back to the orchestrator. Each of those sessions runs the PR through the /adversarial-gauntlet that does its own brief/plan/implementation/post-review each with min 3 rounds of sub-agent review till 0 blocking issues are found.

I only interface with the orchestrator and it handles the session kick off, relays any decisions I need to make, and session clean up.

Been using this for a few weeks now. First run ran for 3 days with 9 PRs. Second round ran for 6 days with 19 PRs. About to run another that will probably take another 5-6 days with 20 PRs.

The biggest issue I had was getting the inter-session messaging to work correctly and not go off the rails messaging the wrong sessions.

But overall it’s been fantastic. I used Fable Extra as the orchestrator and then Opus High for the PRs. Actually seeing better results from Opus High than Extra for the PR sessions. They stick to their lane more and don’t try to branch out to “fix” ancillary issues.

[Andrew Sinagra @sinagra]

@syswarren My wife kept calling Claude 'CC'. I thought it was short for Claude Code, but apparently it is short for 'Captain Chatterbox'.

[Andrew Sinagra @sinagra]

Well this is a new one for Claude:

"The subagent is concurrently editing location/resource — I must not let it observe my sabotage."

Nothing invokes confidence like reading about your agents sabotage and its attempts to hide it.

[Andrew Sinagra @sinagra]

What is odd is that cross session communication has been working great in the desktop app for the last few weeks. 0 issues across multiple orchestrated runs. Ever since an update that was released earlier this week all of a sudden the sessions are having trouble finding other sessions and they are routing messages to the wrong sessions. They also forget which session they are communicating with even after they've established which session is the correct session. For example, an implementation session finds the orchestration session and sends it a message. Then when it needs to report the state of the PR to the orchestrator, it has to re-find the session and will sometimes get it wrong. Additionally, the sessions now have trouble just communicating in general. They will state that the other session is not active or reachable, meanwhile other sessions are actively communicating back and forth with the orchestrator session. I'm not sure what changed but there was definitely some type of regression this week.

[Andrew Sinagra @sinagra]

@amorriscode

[Andrew Sinagra @sinagra]

Congratulations!! Excited to see what you are doing to do. My biggest issues right now are:

- spawned chips need to be able to have their own model and effort set before kicking off
- spawned chips need to survive a restart
- session agents should be able to start a spawned chip session if provided permission
- a backlog of issues from which I can work through. Each would be the start of its own session.

[Andrew Sinagra @sinagra]

Spawned chips need to be able to have their model and effort set before kicking off. Right now it’s bound to the main session so I have to juggle the main sessions settings or start the chip and then immediately stop it to change the model, then resume. Also chips need to survive a restart. Currently they’re lost.

[Andrew Sinagra @sinagra]

@Steve_Yegge nailed it

[Andrew Sinagra @sinagra]

@cjav_dev @Adams_Tweeter @ClaudeDevs I definitely need to play around with using /advisor. How would you compare the quality of running Fable at a lower effort vs using Opus with a Fable advisor? Do you also think running Opus High with a Fable advisor is better than running just Opus Extra?