Building a Fast, Accurate Agent with Prime-RL Post-Training

Ramp Sheets' spreadsheet search agent Fast Ask is a retrieval subagent that finds the numbers or rows the main agent needs on its behalf. In internal traces, 17.8% of the main agent's tool calls were spent opening tabs, reading ranges, and filtering out unnecessary sheets, and about 75% of those immediately led to an additional read call.

So instead of a general-purpose model, Ramp built a specialized model dedicated to retrieval using RL post training. Ramp trained it on top of Prime Intellect's prime-rl stack, Hosted Training, and the verifiers framework, and the result was a roughly 3 billion active parameter model fine-tuned from the Qwen3.5-35B-A3B family. This model delivered latency on par with Claude Haiku 4.5 while scoring 4pp higher than Claude Opus 4.6 on exact-match accuracy in the held-out eval.

The training environment closely mirrored the actual deployment harness.

• It used a financial-style dataset consisting of synthetic workbooks, natural-language questions, and correct answers.
• 14 task types were created, including revenue rollup, invoice reconciliation, spend analysis, time-filtered lookup, and multi-join aggregation.
• Each task was produced in 3 phrasing variants, with an additional invoice-clue variant added for the reconciliation task.
• Batches were composed using a balanced round robin that emitted each of the 14 task types exactly once per pass.
• Search pressure was increased through decoy sheets, partial helper summaries, and identifier obfuscation.
• The only tools available were the 3 tools get_workbook_metadata, read_ranges, and run_python, with a 15-turn budget.

The RL setup was also kept simple. For each question, 8 rollouts were sampled and trained with GRPO using relative rewards within the group; given that tool-use rollouts are slow, Prime Intellect's async off-policy training and an AIPO-style clipped importance-weighted objective were used. Reward prioritized correct ANSWER: parsing and exact match above all, with only a small shaping reward for turn count and unnecessary text. Training ran for 100 steps over roughly 26 hours, with a batch size of 256, evaluated every 20 steps on 128 held-out examples, and thinking mode was turned off to match real low-latency operating conditions.

Training converged quickly from the start. Reward rose from 0.2 to 0.8 over the first 40 steps, and final accuracy improved by 10pp over the base model. Average completion time decreased, but the number of cells read didn't change much, revealing that the policy learned to read better rather than read less.

The core conclusion is clear: for retrieval tasks where the correct answer is deterministic and verifiable, training a specialized subagent separate from the main agent with RL is faster, cheaper, and more accurate.