As infrastructure architects deploying automated coding and sysadmin agents in late 2026, we often treat the agent harness, the ReAct loop wrapper governing planning, tool access, and context compaction, as a static, one-size-fits-all framework. We assume that adding structured planning tools and rich file-editing APIs universally improves output reliability.
A rigorous empirical study testing 176 configurations across Nemotron-3 (30B, 120B, and 550B) and Mistral-Medium-3.5-128B on SWE-Bench Verified (500 repository tasks) and Terminal-Bench 2.1 (89 shell tasks) proves that assumption dead wrong. What acts as essential scaffolding for a smaller model becomes expensive, latency-inducing overhead for a frontier model.
1. The Inversion of Tooling Overhead
When running smaller, cost-effective models (like Nemotron-3 30B), removing structured planning (update_plan) causes two-thirds of runs to abort without making a single code edit; stripping away specialized file tools in favor of a raw terminal causes 66% of shell tasks to crash because the model hallucinates non-existent tool calls.
Conversely, when scaling up to Nemotron-3 550B, removing the 8 specialized file-editing tools and leaving only a raw Bash terminal actually increased accuracy (+3.6 points on SWE-Bench, +5.6 on Terminal-Bench) while slashing operational token costs by 53% and 30% respectively. Instead of making 4.6 fragmented micro-edits per file, the frontier model piped complex shell commands together in single, decisive strikes, cutting total API turns by up to 32%.
2. Two-Threshold Context Compaction (FinOps Optimization)
With narrow context windows (32K tokens), running without context management drops 78.7% of tasks due to overflow. However, relying purely on LLM-generated history summarization is prohibitively expensive because every summary requires a full inference call.
The optimal FinOps pattern is Two-Threshold Hybrid Compaction:
• Soft Threshold (Elision): Replace stale tool outputs with cheap, zero-token placeholders noting the line count stripped.
• Hard Threshold (Summarization): Trigger an expensive LLM summary only when placeholder elision is exhausted, while keeping the initial prompt and most recent turns pinned.
• Skip History Restoration: Giving agents a tool to "restore" elided text proved virtually useless-models ignored it in over half the runs, and when the 30B model did use it, accuracy actually dropped by 3.4 points.
Tune your harness to the exact capability profile of your underlying weights and workload topology.
How are your teams stripping down tool definitions for frontier models?