Coding agents shed frontier calls: apply, search and compaction move to small models

The expensive part of a coding agent is no longer the reasoning call, it is everything wrapped around it. Relace reports that search consumed more than half the tokens across 1,200 agent traces and answers with a dedicated retrieval model at 0.71 Recall@k plus a compaction pass that cut one real trace by 55 percent, while Inception tuned Mercury Edit 2 on logged accept and reject signals to make its suggestions 27 percent more selective. The shared assumption across both is that the frontier model should write new code and do almost nothing else. Teams budgeting agent runs should price that inner loop separately, because it is where the tokens and the seconds actually go. Read: Agentic search buys accuracy with latency because tool calls run one after another. Relace trained a dedicated subagent to fan out parallel view, grep and bash calls, recovering 4x on latency while holding F1 near a frontier model. The motivation is in their own traces: across 1,200 tasks, search consumed over half the tokens spent. Read: For long-running coding agents, input tokens now dominate spend. Compacting context at the cache miss turns cache-read cost from quadratic to linear, measured at 55 percent off one real trace. The constraint is that compaction only pays if it runs fast enough to stay invisible to the agent loop. Read: Inception's reasoning model reaches four-figure tokens per second on standard NVIDIA GPUs by refining a draft in parallel rather than committing token by token. The pitch is that test-time compute stops trading against latency, which is what puts a reasoning model inside a 500ms voice budget that GPT-4.1 currently holds by default. Read: Next-edit models fail by suggesting too much, not by suggesting wrongly. Aligning on logged accept and reject signals with KTO made edits 48 percent more likely to be accepted alongside the selectivity gain, which treats suggestion fatigue as an alignment target rather than a UI problem. Read: An 850M model produces timestamped transcription and translation on a single H100, off query and key embeddings and recovering alignment after generation instead of running a separate forced-alignment pass. It is a bet that timestamped transcription at scale is throughput-bound rather than accuracy-bound. Discuss: Reka's position is that template orchestration, where an LLM is called repeatedly to choose the next step, carries only shallow reasoning, and that decomposition, self-verification and backtracking should be trained into the model instead of supervised at inference time. That is a direct challenge to how most agent harnesses shipping today are built. Read: The breakdown covers performance per megawatt against Blackwell, where the published benchmark comparisons stop holding, Samsung HBM4 supply, and how far AI-assisted design carried the project. It is the same cost pressure driving work off frontier models, read one layer down in the stack.