A task running on Opus 5 or Sonnet 5 now works against the model's full one-million-token window, so a long session keeps going where it used to stop.
The full window. By default, Claude Code checks the conversation against a fixed 200,000 tokens and refuses the next turn itself, before any request leaves the machine. A task that read a large repository for a few hours hit that cap and stopped on a turn the model would have accepted. Opus 5 and Sonnet 5 now run at a million. The other Claude models still run at 200,000, so the model you pick also sets how long a session can run.
Tools load on use. On Claude Code, agents now pull in the tool definitions from your connected MCP servers on demand, so a task with a lot of tools connected keeps far more of its context window for the actual work. An agent starts a turn knowing the name of every tool it can reach, and fetches a tool's full definition before it calls it. On one live session, a single connected server with 73 tools had been putting around 103,000 tokens of definitions into every request.
Platform tools too. Wallfacer's own tool servers, the ones that work on the agent's computer, drive its simulator and send its messages, load on demand the same way. Claude Code's built-in tools that duplicate the ones reaching your code, or have nothing to act on in a Wallfacer session, were removed outright, which frees roughly 13,300 tokens from every request.
Less standing instruction. Reference material an agent needs a few times a session, like the Google Workspace command syntax, has moved out of the standing instructions. The agent looks it up when it needs it.
No prompt size cap. Prompts and messages have no fixed size limit. How much text a task or a message accepts is bounded only by the context window of the model you selected. MCP server configuration keeps its own size limit: it is checked when you save the configuration, and one that is too large is rejected with an error that names the limit.
Start here
Open the agent, go to Configure, expand Advanced settings, and pick Opus 5 or Sonnet 5 under Model. On-demand tool loading is already on for every Claude Code session, with nothing to configure.
Read the models docs for what each model's window is and what happens to a session that outgrows it.