T3 Code · Performance over time

About t3cp.

T3 Code works well for one developer at a laptop. Run it as one server for many coding agents at once, dozens or hundreds of them, and other limits show up: a main thread close to saturation, reads waiting for a database connection, memory that grows by the hour, queued runs that start late.

So we forked it. Our performance fork works on exactly those limits, and it keeps no fix to itself: every change goes back to upstream T3 Code as a pull request, listed on our MRs page. The fork is there to make upstream faster, not to replace it.

Each night we measure the same way. Upstream's newest nightly runs on a cloud machine against the same large database, at 1, 50, 100, 150, 200 and 250 agents, three attempts at each level, and the numbers are published the day they are taken. Each release of the fork is built from its public tag and measured on the same machine, under the same load, against the upstream code it is based on. A change counts only when it is larger than the metric's measured noise; everything else is within noise.

On the charts, upstream's nightlies are one line and our fork another. Each night we build the fork's latest release from its public tag and measure it beside that night's upstream build, on the same machine and under the same load. Where the fork is faster there is a gap between the lines. When upstream merges one of our pull requests, its next nightly carries the fix and the gap at that metric closes. When the lines meet, upstream has every fix the fork had to offer.

How it is measured.

Load
1, 50, 100, 150, 200 and 250 agents. At each level N: N agents, 3 people watching live, switching threads every 300 s, and status polling that grows with N; each agent waits 60 s between turns (5 s at 1 agent). Each agent is one streaming session on the server; 1 agent is one developer starting a new thread on the same large database. At the higher levels the server may saturate: those attempts are published and flagged with how many agents were served, and their latency is context only.
Run
60 s warm-up, 300 s measured, three attempts per level. The value is the median of the three.
Machine
Linux x64: a cloud VM (n2d-standard-16, AMD EPYC 7B13, spot), a new one for every record, its CPU model pinned.
Levels
The overview leads with 1, 50, 100 and 250 agents; every chart page offers all six load levels, and the empty-database cell where it adds information.
Data
A 27 GB synthetic database: 3,000 threads, 1,800,000 turn items. The recipe is published on Methodology and the files on Downloads.
Verdict
Each record is compared with the median of the previous seven. A change counts only beyond the metric's measured noise band plus five points. Everything else is within noise.
Fork
github.com/navidnadali/t3code, built from its public tag on the same machine, and compared with its upstream base: the upstream code at the commit the fork release is based on, built from source the same way.
Empty cell
A seventh cell every day: a fresh install, one agent, on an empty database, for startup, baseline and single-turn metrics. 21 attempts per record.

Not T3 Code.

t3cp.io is not affiliated with, endorsed by or part of T3 Code or its owners. T3 Code is their brand and copyright. This site measures their published releases and links to them.

Run and funded by Corefinity, which pays for the machines and has no other role in the numbers.

Take the data.

CC BY 4.0, attributed to t3cp.io. Every file lists its sha256 on Downloads.

Licences.

Code
MIT, matching T3 Code's own licence (MIT, © 2026 T3 Tools Inc.). Harness code derived from T3 Code keeps that notice; the site, the aggregator and the pipeline carry the project's MIT notice beside it. LICENSE in the repository.
Data
CC BY 4.0, attributed to t3cp.io: "Data: t3cp.io, CC BY 4.0" with a link is enough. data/LICENSE in the repository.

Visits and corrections.

One page-view counter runs: Plausible. It sets no cookie. Corrections: open an issue on the project's public repository on GitHub.