Curiosity Comparison Between Agents

Two studies ago we called it the incurious agent: coding agents that skim the surface of a workspace and skip the substance. We added a third agent to the same test — Google's Gemini — and, for the first time, the pattern breaks. Where one agent greps the surface and another opens almost nothing, Gemini reads. On a lightweight C repository, run head-to-head against OpenAI's Codex and Anthropic's Claude, Gemini opened a real source file in every single run, spent more of its effort reading than either rival, and did the one thing neither of them would: the richer we made the workspace, the more of it Gemini went and looked at.

TL;DR: We put Google's Gemini through the same seven-rung context ladder used in The Incurious Agent, on a lightweight C repository, side by side with Codex and Claude, and measured what each agent actually did. Three findings, and they all point the same way. First, Gemini reads the most: it spent the largest share of its actions — 38% — on reading, against 23% for Claude and 9% for Codex, which puts it at the top of our curiosity index with a score of 100 to Claude's 60 and Codex's 23; it reached for a real file in all seven of its runs, while Codex opened nothing at all in five of its seven. Second, its curiosity scales with the room: Gemini's exploration more than doubled from the empty repo to the full scaffold, while Claude's tapered off and Codex's swung around at random. More context made Gemini look more, not less. Third, the curiosity is engagement, not noise: Gemini was the only agent that reached into the project's own planning docs, and it still landed the core change in every run, two of them cleanly perfect. On the trait the last two studies said was missing, Gemini simply has more of it.

1. A third agent on the ladder

The setup is the one we've used throughout this series, kept deliberately small so the behavior is easy to read. We took a lightweight C repository, pinned it to a single commit, and wrote one well-specified engineering task against it, gated by a check that is red before the work and green after — and re-run by the harness itself rather than trusted from the agent's word.

Then we stood the same seven-rung ladder of context next to it, each rung adding one more layer of standing project material to the same workspace:

• E0 — the bare repository, code only.
• E1 — + README.
• E2 — + AGENTS.md (an operating contract).
• E3 — + PROJECT.md (a project brief).
• E4 — the full scaffold (contract, brief, objectives, plan, progress, a memory tree).
• E5 — the full scaffold plus a seeded memory/ of hand-written notes about this codebase.
• E6 — a leaner "DOX" scaffold with on-demand memory.

The new arrival is Gemini, running through the Antigravity CLI, and we ran it against the two agents from the earlier study — Codex and Claude — under identical conditions. The model, the prompt, the task, and the codebase are held fixed; the only thing that changes down the ladder is how much project context is sitting in the folder when the agent wakes up. Everything below is measured off the run telemetry: which files each agent opened, which tools it called, and what it scored.

2. The agent that opens files

Start with the simplest possible question, the one the whole series turns on: when an agent sits down in a workspace, does it actually open anything? Here is each agent's curiosity index — the share of its actions spent opening files to read, scaled so the most curious agent is 100.

The gap is not subtle. Gemini reads at more than four times Codex's rate, and comfortably more than Claude's — it tops the index at 100, against Claude's 60 and Codex's 23. And the index understates how consistent it is: Gemini reached for a real file read in every one of its seven runs, always cracking open the source before it touched a thing. Claude reads sparingly — a modest, roughly fixed slice of its effort. Codex, tellingly, went through five of its seven runs without opening a single file at all, reconstructing the project from the outside with searches instead of reading it from the inside.

That difference in habit shows up even more sharply when you stop counting files and look at how each agent spends its actions. Bucket every tool call into three kinds — opening a file to read it, probing the surface (grep, shell, git), or editing — and the three agents separate cleanly:

Reading is the single biggest thing Gemini does: 38% of its actions are opening files, more than it spends grepping and more than it spends editing. Claude reads about a quarter of the time and spends nearly half its effort probing the surface. Codex barely reads at all — 9% — and pours 71% of its actions into grep-and-shell, circling the code without ever opening it. One agent reads the room; the other two, to different degrees, feel around its walls.

3. Curiosity that grows with the room

A skeptic could ask whether Gemini just has a fixed appetite for files — that it would read the same amount no matter what's in front of it. It doesn't. Gemini's curiosity is responsive: the more context we put in the workspace, the more of it Gemini went and looked at. Plot each agent's tool activity as you climb the ladder from an empty repo to the full scaffold:

The blue line is the one to watch. In the empty repo Gemini barely moves — there's nothing to explore, so it doesn't. Fill the workspace with the full scaffold and a seeded memory tree, and its activity more than doubles, hitting its peak exactly where there's the most to find. Claude runs the mirror image: it's most active on the bare rungs and quietly does less as the workspace fills — a flat-to-shrinking curve that treats added context as something to get past. Codex zig-zags with no relationship to what's actually in the folder. Only Gemini's effort tracks the environment: give it a richer room, and it does more looking. That is the exact shape of the trait the last two studies kept finding absent — an agent that responds to what you put in front of it.

4. Curiosity you can see in the work

None of this would matter if the reading were idle. It isn't. Gemini's curiosity shows up as genuine engagement with the workspace — and then in the result.

On the richer rungs, Gemini was the only agent that opened and updated the project's own planning document, ticking off the task in the scaffold's checklist as it went. The other two never touched it. You can read that two ways, but both are flattering: Gemini engaged with the project's structure as a living thing to work with, not decoration to step around. It was curious enough about the scaffold to use it.

And the reading paid its way. Gemini got the core change right in every single run, two of them scoring cleanly perfect, the rest landing the exact code change asked for. Where Gemini shed a fraction of a point, it was never for the logic — it was for being thorough, touching a file or two beyond the strict minimum, the same engagement that had it updating the plan. Codex, by contrast — the agent that barely opened anything — never once satisfied the gate across all seven runs. The agent that read the most produced the work that worked; the agent that read the least didn't. Curiosity wasn't a tax here. It was the difference between looking and guessing.

5. Curiosity, finally

Put the three charts together and one trait explains all of them.

Environmental curiosity — an agent's tendency to seek out the parts of its environment that would help it, in proportion to how much they would help, rather than reading whatever happens to be in front of it, or nothing at all.

The Incurious Agent named this trait by its absence: Codex read the signposts but never walked the path, Claude didn't read the map at all. Gemini is the first agent in this program that starts to look like it has it. It reads more of its workspace than either rival, it spends more of itself reading than doing anything else, and — the part that matters most — it reads more when there's more worth reading. On the exact axis the series has been circling, Gemini is the agent that moved.

6. How to read this honestly

This is a small, controlled probe, and the caveats shape the claim.

• A single run per rung, one small repository, one task. These are directional behaviors, not tight confidence intervals — enough to see a clean, repeated separation in reading habits, not enough to rank the fine details. The effect we're reporting is the large one: which agent opens files and which doesn't.
• The deepest context stayed unopened by all three. No agent in this run cracked the seeded memory/ tree specifically; the curiosity we measured is Gemini opening source and project files far more than its rivals, not evidence that any agent yet mines curated memory. That deeper frontier is still open, and it's the next thing to push on.
• Thoroughness has a price. Reading costs tokens, and Gemini's habit of opening files makes it dearer than an agent that greps and leaves. We think that's the right trade — the frugal agent here is the one that failed the gate — but it is a trade, and on a much larger repository a read-heavy agent is where cost would start to bite.

The harness, the ladder, and every run are saved and reusable. The next round raises the repository size and the task difficulty until curiosity has more to win — and, on this evidence, points it at the agent most likely to win it.

---

Appendix: per-agent, per-environment data

Every number above is recomputed from the per-run telemetry — reading share counted from each run's event log, scores from the harness gate. One run per cell.

"Read %" is the share of an agent's tool calls spent opening files to read. The curiosity index (§2) rescales each agent's mean read share so the most curious agent is 100: Gemini 100, Claude 60, Codex 23. Exploration (§3) plots each run's activity as a share of the busiest run.