Choose an AI backend, and fail over when you hit a usage limit

Every agent in MyAgens runs on a backend, and the backend is the swappable part. Your agent's persona, its memory, its skills, its connectors, its approval rules, and its cards all live in MyAgens, not in the model, so pointing an agent at a different backend does not cost it anything it knows.

That matters for two everyday reasons: you can put cheap or local models on routine work and keep a frontier model for the hard parts, and you can keep working when one provider tells you to come back tomorrow.

The six backends

Each backend wraps that provider's own agentic CLI, with its tool belt, sandboxing, and permission modes, rather than reimplementing a tool loop. Claude is the default everywhere.

Each backend brings its own account and its own sign-in. The CLI handles that itself, so MyAgens is not holding a second set of API keys for you.

What you give up, honestly

Claude is the most complete backend. The others are close, and worth knowing where they are not:

• Codex, Antigravity, and Cursor are full MyAgens agents. They get your tools (memory, the board, skills, crew messaging, and every enabled connector, under the same names Claude uses, so an always-allow preset you granted on Claude means the same thing here), your agent (persona, work.md, known directories, the crew roster, and the memories recalled for that message), and your approvals (below Full autonomy, risky calls stop for Approve or Deny, and a refusal reaches the model as a failed call with your reason).
• Cursor reports live text, tool calls, and real token usage, so tool activity and per-agent cost tracking work the way they do on Claude.
• Antigravity reports no token or cost accounting, because its CLI does not emit any. The work happens, the spend just is not counted.
• Grok streams text and resumes sessions but shows no tool-call visibility or usage data.
• Connectors that run as their own process (Unreal Engine, Unity, and the Browser Sketchpad) are Claude-only. Antigravity's own browser tools cover the last one for that backend.
• Ollama is plain chat, not an agentic CLI. It is there so an agent can run fast and completely independent of any cloud provider on a small local model. Do not expect it to drive multi-step tool use.

Autonomy behaves differently on the CLI backends. agy and cursor-agent auto-approve their own tools when run non-interactively. Below Full autonomy MyAgens confines the run inside that CLI's own sandbox (agy --sandbox restricts the terminal to the workspace), and Full autonomy lifts the gates entirely. Read Approvals and autonomy before you put a non-Claude backend on Full.

Switch an agent's backend

From a chat, on the agent you are talking to:

/model agy-cli
/model cursor-cli:gpt-5
/model claude-agent-sdk

The <backend>:<model> form picks a specific model on that backend. Leave the model off and the backend uses its own default.

From the panel, use the AI backend selector in Settings (for Atlas) or in a worker's form (for a Lead or assistant). For Antigravity and Cursor the Model field has a Fetch button that lists the exact model labels your installed CLI accepts, so you are picking from real options rather than typing an id and hoping.

Every agent chooses independently. Atlas can sit on Claude while your DevOps lead runs on Codex and a research assistant runs on a local model.

Switching an agent's backend resets its stored session id, because a resume handle from one provider's CLI means nothing to another. The next turn starts a fresh conversation on the new backend. Everything in memory is unaffected.

Set a fallback for usage limits

Any agent can carry a fallback backend and model of its own, and every Lead and worker fails over independently of Atlas. Configure it next to the primary in the same place: Settings for Atlas, the worker's own settings for anyone else.

Two different things trigger it:

1. Proactively, before the wall. When the cached usage probe shows your plan is at or over its rate-limit threshold, background and autonomous turns move to the fallback on their own. Interactive turns are deliberately left alone, so a chat you are sitting in front of does not silently change model under you.
2. Reactively, mid-flight. When a turn hits a usage or rate-limit error while it is running, interactive turns included, MyAgens retries that turn once on the fallback and tells you in chat where it landed.

The retry is one attempt. A second failure propagates as a normal error rather than cascading through backends.

A fallback can be a completely different backend, not just a second Claude endpoint. A practical setup is a frontier primary with a local Ollama fallback, so the worst case is a slower answer rather than no answer.

A cross-backend failover starts a fresh conversation, for the same reason a manual switch does. MyAgens only carries a resume token within one backend. Failing over mid-task means the retry does not remember the earlier turns of that conversation, so it works best on self-contained work.

Not every error is a usage limit. Authentication failures, the stall watchdog, and a turn you cancelled yourself are deliberately excluded, because failing over would not fix any of them.

Which one should you actually use

• Claude unless you have a reason otherwise. It is the default, the most complete, and the only one with every connector.
• Codex or Cursor for a second full-strength agent on a separate account and a separate rate limit. Cursor if you want the usage numbers.
• Antigravity for the Gemini tiers, if you can live without cost accounting.
• Ollama for a fallback that cannot be rate-limited, or for work that must never leave the machine.
• Grok if that is the account you have.

Related

• Bring your own model, run agents locally or on any provider covers pointing MyAgens at a local or OpenAI-compatible endpoint.
• Do I need an API key, or does my Claude subscription work? for the Claude side specifically.
• Build your crew, Leads and specialist agents for creating the agents you are assigning backends to.