Specifically, I'm seeing inaccurate agent code generation (since it's using "dumb" tools like shell commands vs. smart ones like MCP) which is also token expensive. Moreover since we're not doing anything to scope it, I'm seeing context bloat as token use compounds with context scaling multiplicatively over time. We should be able to improve this with some agent/harness configuration.
Specifically, I'm seeing inaccurate agent code generation (since it's using "dumb" tools like shell commands vs. smart ones like MCP) which is also token expensive. Moreover since we're not doing anything to scope it, I'm seeing context bloat as token use compounds with context scaling multiplicatively over time. We should be able to improve this with some agent/harness configuration.