中文
AI Engineer World's Fair

Replacing 15 tools with a single REPL: the full post-mortem of taking a spreadsheet agent from 50% to 92%

Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs · Nuno Campos

19 min
AgentAI CodingEvalsContext

19 min total·Actually worth watching closely: ~4 min·2 must-watch clips

Orange = the 4 minutes worth watchingFor the rest, the guide is enough
Segment guide · 7 segments
  1. 0:17 2:10Listen

    An underrated hard problem: why LLMs can't handle spreadsheets

    Opens by framing the problem: humans see a spreadsheet's structure instantly and visually, while an LLM has to disambiguate layout and meaning on its own, which makes spreadsheets a blind spot for coding agents.

    How hard spreadsheets are for LLMs is systematically underrated, and the hard part is structural disambiguation, not the calculation

    Purely spoken setup of the problem with nothing on screen to depend on, fine to listen to like a podcast▶ Jump to 0:17
    Speaker · Nuno Campos
  2. 2:10 5:00Listen

    The dead-end list: multi-agent and SQL/XML representations all failed

    A post-mortem of the failed paths: the multi-agent architecture, SQL, XML and every other single representation. The exploration did leave two byproducts behind, CSV/TSV partial views and an HTML-inspired render engine.

    The failed representations weren't wasted; what remained of them became parts of the final architecture

    Mostly narrative post-mortem, with nothing in this chapter worth watching, so just listen to why each one failed▶ Jump to 2:10
    Speaker · Nuno Campos
  3. 5:00 7:20Listen

    The biggest breakthrough: one REPL with persistent state

    Replacing about 15 tools with a single Node.js REPL: variables survive across calls, one call combines several operations, and the agent can interleave reasoning between steps, where before it took 10 or 15 sequential calls that often timed out.

    If your agent is making many sequential tool calls, you're really inventing a bad scripting language

    The core argument is delivered verbally with no demo on screen, but it's the densest stretch of the talk and deserves full attention▶ Jump to 5:00
    Speaker · Nuno Campos
  4. 7:20 10:50Listen

    Building your own verification loop: formula engine plus render engine

    Following the pattern where coding agents iterate against a compiler and tests, the team built a formula engine and a render engine as the source of truth driving self-correction, with a warning that a low-fidelity engine, one implementing only half of Excel's formulas, is worse than no loop at all.

    The loop is only as good as the engines that power it, and a half-built source of truth turns right answers into wrong ones

    Engineering decisions and pitfalls with no live demo, so focus on following the argument for why building it yourself is worth it▶ Jump to 7:20
    Speaker · Nuno Campos
  5. 10:50 14:22Listen

    Interfaces go stale, verification loops don't

    Separating the two assets: the REPL is the best interface today but may change as the labs push computer use, while four or five model releases showed that more capable models get more out of the verification loop, making the engines the durable asset.

    Invest separately in the interface that will go stale and the loop that won't

    Strategic reflection delivered purely in speech, good to listen to while mapping it onto how your own project's assets are layered▶ Jump to 10:50
    Speaker · Nuno Campos
  6. 14:22 16:40Skim

    Deterministic evaluation and bugs disguised as reasoning failures

    Replacing LLM grading with black-box comparison against a golden spreadsheet's inputs and outputs, plus the point that a lot of what looks like the model being dumb is really infrastructure bugs: a wrong example in the prompt followed faithfully, or failing tools that trigger endless retries.

    Go deterministic instead of an LLM judge wherever you can, and read the traces to tell 'the model didn't get it' from 'something we can fix'

    The benchmark numbers (50%→92%, the jump to 74%) are most likely on slides, so a glance at the data slide alongside listening is enough▶ Jump to 14:22
    Speaker · Nuno Campos
  7. 16:40 18:48Listen

    Six generalizable lessons for building agents

    The wrap-up: the REPL interface, building your own feedback loop, deterministic evaluation, and domain knowledge narrowing the model's attention, with the emphasis that domain knowledge isn't teaching the model concepts but pigeonholing an all-knowing model onto what the current task needs.

    Domain prompts kept improving results across every iteration of the tools and port almost verbatim between approaches

    The closing summary depends on nothing on screen, is the best part of the talk to take notes on, and is worth a second listen▶ Jump to 16:40
    Speaker · Nuno Campos