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
Congrats on the launch. The routing question is already well covered above, so a different one.
The part I would poke is memory over long unattended runs. "Remembers your whole codebase" plus "keeps working in your off-hours" is the exact case where drift bites. On a long run the working context gets summarised to stay affordable, and a summary of the codebase slowly stops matching the codebase. Nobody is watching, so the agent keeps building confidently against a picture that has gone stale.
So my question, does Codi re-read the actual files as ground truth during a run, or does it work from its remembered version of them? Because for a supervisor running for hours with no human in the loop, what it remembers about the code and what the code actually says drift apart, and that gap is invisible until the mroning.
Not a gotcha, it is the thing I get wrong most on long agent runs myself.
AskCodi
@jernej_jan_kocica by actually dividing the whole system into supervisor and worker, we are eliminating long runs. The entire project is divided into features and modules. Agents are fired our features -> agent comes back with a PR -> supervisor reviews and merges. After every 4-5 PRs, board and codebase is re-synced so that supervisor doesn't go off plan.
@shreyans_assistiv Makes sense, short worker runs can't drift far before a reviewable PR lands, and the 4-5 PR re-sync stops the supervisor going off-map. The part I'd still watch is intent, not code. When this has broken on me it wasn't the supervisor forgetting the codebase, it was forgetting why a call was made a few PRs back and then approving one that quietly undoes it. Each worker only sees its own module, so cross-module drift is invisible to everyone but the supervisor, and that's the one component running from a summary. When you re-sync, do you pull the original spec and acceptance criteria back into the supervisor's context, or mainly the current board and code state?
Honestly
congratulations on the launch! What is the differentiation here between AskCodi and me instructing my agents to run in parallel while using the most optimized models for a given task in Cursor or Conductor? In my mind I just create a skill about this but I could be missing something.
AskCodi
@scott_davidson_jr It would be still per project, you would have to still manage atleast one agent per project. Fine at one, starts causing issues as you start working on more. This is not just multi tasking, this is a freelancer on steroids. Plus we have built in systems which actually remove you as a bottle without going off rails. Of course you could build out this system will skills, but skills are still suggestions for agents in long loops.
Do you currently run a multi agent auto model system in Cursor or Conductor? Last time I saw someone's stats, autorouting was still 98% opus. Would love to learn more.
@shreyans_assistiv congrats on the launch. Which models do you find are being used most often?
AskCodi
@adam_maceachern1 Kimi K3, GLM 5.2 and Sonnet 5 as the more effective worker models. Fable or Sol for supervisor. The system is absolutely killing it :) Thank you for your support!!
the overnight autonomy part is what I keep coming back to. if Codi picks a wrong approach at 2am and just keeps going since nothing hit the "needs a decision" bar, do you find out in the morning after it burned hours down a bad path, or is there some checkpoint that catches a run going sideways before it eats your whole night's budget on the wrong thing
MakersClaw
@omri_ben_shoham1 Codi works in bounded tasks, not one open-ended overnight run. Each task has acceptance criteria, tests, and budget limits. If validation repeatedly fails or the work drifts from the project charter, Codi pauses that branch and surfaces it in the inbox instead of continuing down the wrong path.
The inbox model for agent blockers is a smart way to cut the noise compared to watching a chat thread spiral. Curious how AskCodi scopes context between projects — does each project get its own isolated context window, or do the agents share ambient workspace state across the whole account? Also wondering if there is a mode where sensitive code stays local and only higher-level summaries go to the cloud.
AskCodi
@hi_i_am_mimo each project gets it own isolated context window, even task is completed in their own worktrees. Codi is the only one who can see the status across projects, tasks and agents. We have context filtering and PII masking already implemented so the agents avoid sensitive code before it hits the agent loop.
Bond
interesting!! where's the name from?
AskCodi
@teaganyuen Its actually a 4 year old brand name, it changed from AskJarvis. We always envisioned Codi as a CTO, now it is technically possible.
@Sachin Sharma that answers it well, pausing and surfacing in the inbox instead of plowing ahead is the right failure mode. one follow-up: when validation keeps failing on the same branch, does Codi ever try a different approach on its own before giving up, or is a repeated failure always a hard stop that waits for you to redirect it?
AskCodi
@omri_ben_shoham1 chiming in here, codi tries different approaches, different skills, different models before surfacing it as a blocker for humans.
The "ran your off-hours, we stopped hedging" point is the interesting one. Once Codi runs unattended overnight across several projects, what happens when one of those runs goes wrong, like a bad migration or a broken deploy, before a human sees it in the morning? Is there a rollback path per project, or does the human review catch it after the fact the next morning?
AskCodi
@rahulladumor all the work done is in worktrees, nothing hits prod without human supervision. Everything is supervised by codi plus it follows TDD. Systems are in place to exactly stop this.
the /standup mode starting sessions "based on need and token usage left" is the detail I keep coming back to. when the budget is getting tight and you've got three projects going at once, who's actually making the call on which one gets the remaining tokens - do you set priority ahead of time, or is that Codi's own judgment in the moment based on what it thinks matters most?
AskCodi
@galdayan Standup sessions are exactly how it would work in real work, codi and you decide together the priority list. Codi surfaces what it thinks matter most for your KPIs and then you agree/disagree/modify :) We are continuously experimenting with it
@shreyans_assistiv good to know it's a two-way call, not codi just deciding alone. does that agree/disagree/modify step happen once when you kick off the standup, or can codi come back mid-run and ask you to re-arbitrate if it hits a fork partway through a project? asking because with 3 projects going at once, i'd want to know if that's a one-time setup decision or something that pings me repeatedly during the day.
Prefactor
What models is it run on?