### Problem
We want to start continuously tracking our progress on agent evals over
time. As part of this, we'd like the *score* to have a clear,
interpretable meaning. Right now, it's a number from 0 to 5, but it's
not clear what any particular number works. In addition, scores vary
widely from run to run, because the agent's output is deterministic. We
try to stabilize the score using a panel of judges, but the behavior of
the agent itself varies much more widely than the judges' scores for a
given run.
### Solution
* **explicit meanings of scores** - In this PR, we're prescribing the
diff and thread criteria files so that they *must* be unordered lists of
assertions. For both the thread and the diff, rather than providing an
abstract score, the judge's task is simply to count how many of these
assertions are satisfied. A percentage score can be derived from this
number, divided by the total number of assertions.
* **repetitions** - Rather than running each example once, and judging
it N times, we'll **run** the example N times. Right now, I'm just
judging the output once per run, because I believe that with these more
clear scoring criteria, the main source of non-determinism will be the
*agent's* behavior, not the judge's
### Questions
* **accounting for diagnostic errors** - Previously, the judge was asked
to incorporate diagnostics into their abstract scores. Now that the
"score" is determined directly from the criteria, the diagnostic will
not be captured in the score. How should the diagnostics be accounted
for in the eval? One thought is - let's simply count and report the
number of errors remaining after the agent finishes, as a separate field
of the run (along with diff score and thread score). We could consider
normalizing it using the total lines of added code (like errors per 100
lines of code added) in order to give it some semblance of stability
between examples.
* **repetitions** - How many repetitions should we run on CI? Each
repetition takes significant time, but I think running more than one
repetition will make the scores significantly less volatile.
### Todo
* [x] Fix `--concurrency` implementation so that only N tasks are
spawned
* [x] Support `--repetitions` efficiently (re-using the same worktree)
* [x] Restructure judge prompts to count passing criteria, not compute
abstract score
* [x] Report total number of diagnostics in some way
* [x] Format output nicely
Release Notes:
- N/A *or* Added/Fixed/Improved ...
---------
Co-authored-by: Antonio Scandurra <me@as-cii.com>
The old one wasn't linking, and
https://github.com/zed-industries/zed/pull/29081 has a bunch of merge
conflicts. Wanted to start simple/small.
## Todo
* [x] Remove low-signal examples
* [x] Make the eval run on a cron, on main, and on any PR with the
`run-eval` label
* [x] Noise in logs about failure to write settings
```
[2025-04-21T20:45:04Z ERROR settings] Failed to write settings to file
"/home/runner/.config/zed/settings.json"
Caused by:
No such file or directory (os error 2) at path
"/home/runner/.config/zed/.tmpLewFEs"
```
* [x] `Agentic loop stalled`
(https://github.com/zed-industries/zed/actions/runs/14581044243/job/40897622894)
* [x] Make sure that events are recorded in snowflake
* [ ] Change judge criteria to be more explicit about meanings of scores
Release Notes:
- N/A
---------
Co-authored-by: Antonio Scandurra <me@as-cii.com>
Co-authored-by: Agus Zubiaga <hi@aguz.me>
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Thomas Mickley-Doyle <tmickleydoyle@gmail.com>
This pull request will print all the used tools and their failure rates.
The objective goal should be to minimize that failure rate.
@tmickleydoyle: this also changes the telemetry event to report
`tool_metrics` as opposed to `tool_use_counts`. Ideally I'd love to be
able to plot failure rates by tool and hopefully see that percentage go
down. Can we do that with the data we're tracking with this pull
request?
Release Notes:
- N/A
Now that we've established a proper eval in tree, this PR is reboots of
our agent loop back to a set of minimal tools and simpler prompts. We
should aim to get this branch feeling subjectively competitive with
what's on main and then merge it, and build from there.
Let's invest in our eval and use it to drive better performance of the
agent loop. How you can help: Pick an example, and then make the outcome
faster or better. It's fine to even use your own subjective judgment, as
our evaluation criteria likely need tuning as well at this point. Focus
on making the agent work better in your own subjective experience first.
Let's focus on simple/practical improvements to make this thing work
better, then determine how we can craft our judgment criteria to lock
those improvements in.
Release Notes:
- N/A
---------
Co-authored-by: Max <max@zed.dev>
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Agus <agus@zed.dev>
Co-authored-by: Richard <richard@zed.dev>
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Antonio Scandurra <me@as-cii.com>
Co-authored-by: Michael Sloan <mgsloan@gmail.com>
The `always_allow_tool_actions` setting would get overridden with the
default when we loaded each example project, leading to examples
stalling when they run a tool that needed confirmation. There's now a
separate `runner_settings.json` file where we can configure the
environment for the eval.
Release Notes:
- N/A
---------
Co-authored-by: Oleksiy <oleksiy@zed.dev>
Release Notes:
- Fixed a regression that caused the agent to hang sometimes.
---------
Co-authored-by: Thomas Mickley-Doyle <tmickleydoyle@gmail.com>
Co-authored-by: Nathan Sobo <nathan@zed.dev>
Co-authored-by: Michael Sloan <mgsloan@gmail.com>
Also, we moved the logic for driving the agentic loop into `Thread` so
that we don't have to re-implement it.
Release Notes:
- N/A
---------
Co-authored-by: Nathan Sobo <nathan@zed.dev>
Michael Sloan
Thomas Mickley-Doyle
