the maddening part is every client redoing the same oauth dance. my coding agent already did it, so i stopped authing anything else. my scripts just read the token the agent stored. one auth,…

17 is the right default for chat. for coding agents i skip the tradeoff entirely: the agent writes one script against the api instead of chaining tool calls. measured one task both ways: 40 calls put…

nice, toolCallers: ['code'] is the right default. same idea one layer out: the coding agent is already logged into its MCP servers, so it writes one script against them instead of 40 tool calls.…

either name undersells it a bit. measured one task both ways: 40 MCP tool calls = 262,159 chars in context vs one script the agent wrote = 903 chars. and the script keeps running from cron with no…

agree on 2. the tool call part of that noise you can remove instead of compact. measured one task both ways: 40 MCP tool calls = 262,159 chars in context, re-read every turn after. same task as one…

the square is the part that hurts, and tool results are most of it. they sit in context for every turn after. measured one task both ways: 40 MCP tool calls = 262,159 chars in context, re-read every…

the fastest MCP call is the one no model makes. 65k tokens of Linear tickets through the model, or 225 through a script your agent writes. same answer, no API keys, already authenticated. introducing…