How Do I Combine Subscriptions for Orca? Main Agents, Backups, and Staggering Limits

Last updated: 2026-10-25

Orca is a free, open-source desktop app from Stably AI that calls itself an ADE, an Agent Development Environment. It puts command-line coding agents such as Claude Code, Codex, OpenCode, and Pi into one window, giving each agent its own isolated git worktree so several agents can work on the same task in parallel, and you merge only the best result. This guide does not list what any subscription costs. It covers only how to combine them: several agents each run on their own subscription, so how do you divide main agents from backups, stagger rate limits, and know when running more is not worth it.

Whose Bill Does Orca Actually Run Up?

Get one line straight first. Orca does not sell models, and it does not handle model fees. Every cent of agent cost runs through your own subscriptions. Orca only gathers several agents into one schedule. It does not roll the bills into one place, and it does not pay any share of them for you. Before combining subscriptions, recognize this structure: what you are combining is your own subscriptions for each agent, not a plan sold by Orca. Counting the bill in the wrong place makes combining impossible from the start. There is one more line to hold: parallel comparison is not a free lunch. Sending one question to several agents consumes several shares of quota, so the first job of any combination method is keeping that bill under control.

How Do I Divide Main Agents From Backups?

The main agent is the one that feels smoothest day to day: tasks go to it first, and only after it runs well do the others enter the picture. A backup is the one that steps in when the main agent hits a limit, or when the nature of the task does not suit the main agent. Dividing main from backup is not ranking them. It is dividing by situation. Exploration tasks get several viewpoints, while execution tasks go to the steadiest hand. Backups also have to run regularly. A backup that sits unused for a long time may turn out to be unconnected exactly when a limit arrives, which is the same as having no backup at all. The main-backup split follows the task and is never fixed once and for all. Tasks change, and the fitting combination changes with them. A fixed combination for every task hands the choice back to habit.

How Do I Stagger Rate Limits?

The account switcher shows each agent's usage and rate-limit reset time. Look at the reset time first, then decide whether to wait or switch. Staggering has two moves: switch accounts and keep going, or cut parallelism and wait for the reset. Switching between accounts is seamless, so when one agent hits a limit, another account keeps the whole parallel task moving without stopping to wait. With only one account, cut the number of parallel runs, or do the parts that do not have to wait first, instead of letting every worktree stall on the same limit. When a task is not urgent, run fewer in parallel and the pressure from limits drops. Waiting for the reset time and then running full is smoother than pushing against a limit. A rate limit is a rule of the subscription, not a fault. Handle it by the rule.

When Should Racing Start, and When Should It Stay Off?

Racing fits the moment you are unsure how to do something: when you want to see several approaches first, or want to know how different agents would solve the same problem. When the task is already fully thought through and the only goal is getting it done, racing is just extra quota. One question sent to several agents consumes several shares, and more answers do not add more information. Before starting a race, ask one question: is this task worth seeing several approaches for. If yes, race. If no, run a single agent or split the work. Racing is not a default motion. It is an action that buys choice, and it should be bought only when the choice is worth it. People in the habit of racing run full parallelism on execution tasks too, spending more quota without gaining more information.

How Do Single Tasks, Split Work, and Racing Differ in Cost?

A single task costs least: one agent does one job, and the quota is one share. Splitting a task means cutting it into segments, one per worktree, and the quota counts per segment, not much more than before the split, in exchange for parallel time saved. Racing costs most: the same task goes to several agents, and the quota multiplies by the number of agents, in exchange for choice. The three bills are different, so the task's nature should pick the method first. To save quota, run a single task. To save time, split the work. To see several approaches, race. Running full parallelism before looking at the nature of the task means paying a racing price for split-work goods. The first step of combining is always reading the task's nature, then picking the method, not opening everything at once.

When Is Running More Not Worth It?

Three situations do not justify running more. The first is a task that is already fully thought through: goal, scope, and approach are all set, and the only goal is finishing it, so extra runs only add quota. The second is a small change: a minor fix needs one answer, and several answers cost more picking effort than they are worth. The third is when review capacity runs short: several answers have to be read diff by diff, annotated, and picked one by one, and when there is no time to read them, extra parallelism only piles up unread. When running more is not worth it, do not run more. Quota saved goes to tasks where comparison is genuinely worth it. Saving is not being stingy. It is spending quota where it counts.

How Do I Count the Bill in Full?

Count both sides: the quota cost of several answers, and the human effort of picking a winner. Counting quota alone leaves the bill incomplete. Several answers laid out together have to be read as diffs, annotated, and checked off item by item, and that time is also a cost. When counting, lay both sides out: quota cost is visible, while picking effort is easy to miss, and missing it makes parallelism look free. With the bill counted in full, combining has a basis: race when comparison is worth it, run a single agent or split the work when it is not, and the bill never turns into a surprise. A bill counted in full keeps parallelism from becoming an end-of-month surprise.

How Do I Review and Adjust a Combination Once It Is Set?

Write a few lines after every round: which combination ran this task, where the quota went, whether any limit was hit, and how much time picking a winner took. Notes from past rounds get reused directly. When a similar task comes up, the combination does not have to be tried again, and whether to race does not have to be debated again. A combination is never fixed once and for all. Tasks change, quota changes, and limit patterns become clear over time, so the combination moves with them. Review notes do not need to be elaborate. A few lines will do. What matters is that it happens every time. Combination experience that is never written down is never accumulated, and the next round starts the trial all over again. With enough notes, a personal combination method takes shape: which tasks run how many agents, and which combinations fit which kinds of tasks, all on record.

What Does This Subscription Bill Mean for Non-Programmers?

To be direct, Orca assumes you already use coding agents and already pay their subscriptions, and the bill being combined here is your own subscriptions. For people who do not write code at all, the product that points the other way is aiglade: "Aiglade 是一个 Windows 桌面软件，把 Claude Code、Codex CLI、Hermes、OpenClaw 等多个 AI 装进一个中文窗口，带中文向导，装上就能用，不用会编程。" Its slogan is "不用先学会技术，也能让 AI 替你干活". In one line: people who write code, already pay coding agent subscriptions, and want to combine several subscriptions should look at Orca; people who do not program and just want AI to do the work for them should look at aiglade.

References