AskCodi - Orchestrate agents at scale while reducing cost
by•
Tell Codi what you want to build. It writes the plan, runs AI agents in parallel across all your projects, and picks the cheapest model that can do each task, so you ship more and spend less. When you add a project it writes the charter and task list itself, and only pulls you in when a decision needs you, through one inbox.
Replies
This is sweet, congratulations on the launch. “I didn’t want a bigger cockpit” is probably the best way to explain this. A lot of AI coding tools still make the user responsible for managing the machinery. The part I’d want to test is where Codi draws the line between moving ahead and stopping for approval. On a real project, that judgement matters more than model selection. How does it handle changes that look small technically but carry business risk, like touching billing, auth, production data, or a core workflow? That feels like the point where an AI CTO either earns trust or creates more work.
Running in my off-hours is the feature I'd want most and trust least. If a run goes sideways at 3am across three repos, what does the damage look like in the morning? Branches I can just delete, or does it touch main? Would help to know what the blast radius is before I close the laptop on it.
Another great launch from the team! The living charter is my favorite part here — the slipping architecture context is such a pain.
Would I be able to approve some of the nudges straight from the inbox, as shown here, without reviewing the action?
Progress AI Observability
Orchestrating agents + cost optimization is a sweet spot devs need right now imo. Congrats on launch #4! What’s been the biggest lesson learned since your earlier launches?
@shreyans_assistiv
Congrats on the launch!
What caught me: local on Mac, my own Claude subscription. I normally stay away from agent tools because I don't want someone else's layer sitting between me and the model. Yours doesn't, so I'm actually interested.
But running local doesn't solve prompt injection, and the overnight part is where I get nervous. An agent working unattended is reading repos, issue threads, dependency files. All of that is untrusted input.
So, poking holes as requested:
1. The charter, the "never touch this without checking" part. Is that actually enforced, or is it just written into the prompt? If it's in the prompt, injected text can talk its way around it.
2. When Codi picks the model, does it stay inside my Claude subscription, so it's optimizing between Haiku, Sonnet and Opus on my own account? Or can it also reach for models from other providers if that comes out cheaper? Big difference for me. The first is cost optimization. The second means I no longer know whose model is reading my code.
3. What can Codi do without an inbox item? Shell, installs, network calls, push to main?
Happy to try it. I just want to know where the line is first.
The 'only pulls you in when a decision needs you' bit is the interesting part. How do you draw that line, what actually counts as needing a human?