Ceiling on a tool's description — an OPERATING POINT, not a provider limit.
No tool-capable vendor enforces a cap anywhere near here: OpenAI (gpt-5.6-luna), Cerebras
(gpt-oss-120b) and Gemini (3.7 Flash) each accepted an 8,000-character description on a live
call. The number is chosen from measurement instead. A comparison run used the
same 1,026-observation corpus at 1,024 and at 4,096 (results doc §10–§11): at 1,024 — where a
large Action degrades to output names and result codes with descriptions trimmed — native
selection accuracy improved on two of three models (+6.3pp Gemini, +11.3pp GPT 5.6-luna). At
4,096, with every description in full, both gains disappeared, GPT-OSS-120B fell a further 5pp,
its prompt grew by ~700 tokens per call, and it began returning no action at all on cases it
had handled. Same-case comparison: −10 correct turns net.
So the names and codes are what discriminates; the sentences around them are dilution the
smallest model pays for most. One A/B at N=3 is directional rather than definitive (envelope-arm
run-to-run noise is ±2–3pp), and a per-provider ceiling is the obvious follow-up — but until
someone measures that, compact is the point the data favours. The detail-degradation below is
therefore the NORMAL path for large Actions, not a safety net.
Ceiling on a tool's
description— an OPERATING POINT, not a provider limit.No tool-capable vendor enforces a cap anywhere near here: OpenAI (gpt-5.6-luna), Cerebras (gpt-oss-120b) and Gemini (3.7 Flash) each accepted an 8,000-character description on a live call. The number is chosen from measurement instead. A comparison run used the same 1,026-observation corpus at 1,024 and at 4,096 (results doc §10–§11): at 1,024 — where a large Action degrades to output names and result codes with descriptions trimmed — native selection accuracy improved on two of three models (+6.3pp Gemini, +11.3pp GPT 5.6-luna). At 4,096, with every description in full, both gains disappeared, GPT-OSS-120B fell a further 5pp, its prompt grew by ~700 tokens per call, and it began returning no action at all on cases it had handled. Same-case comparison: −10 correct turns net.
So the names and codes are what discriminates; the sentences around them are dilution the smallest model pays for most. One A/B at N=3 is directional rather than definitive (envelope-arm run-to-run noise is ±2–3pp), and a per-provider ceiling is the obvious follow-up — but until someone measures that, compact is the point the data favours. The detail-degradation below is therefore the NORMAL path for large Actions, not a safety net.