We Cut Our CI Feedback Time by 39% with Blacksmith#

Our CI had become maddeningly slow. The tests were only part of the problem...jobs routinely spent another minute or two waiting for GitHub-hosted capacity.

We recently moved our critical workflows to Blacksmith in kagent PR #2659 (https://github.com/kagent-dev/kagent/pull/2659), the difference was immediate.

The Results Speak For Themselves#

Across successful main-branch runs, median CI duration dropped from 23 minutes 16 seconds to 14 minutes 10 seconds: a 39% reduction.

That gives developers feedback roughly nine minutes sooner on every change.

A near-controlled comparison—using the migration PR and its parent, which differed only in CI configuration—showed the same result:

That run fell from 19m27s to 13m33s, a 30% reduction.

Jobs now start when they are ready#

Runner pickup was arguably the most satisfying improvement.

We measured the delay between our setup job completing and dependent build or test jobs starting. Median pickup latency dropped from 82 seconds to 16 seconds—an 80% reduction.

Previously, our CI fan-out routinely sat idle for another one or two minutes. With Blacksmith, those jobs generally start within seconds.

The workloads on our critical path also ran faster:

Job	Before	Blacksmith	Reduction
UI tests	9m57s	7m31s	24.5%
End-to-end tests	16m21s	12m22s	24.3%
Runner pickup	1m22s	16s	80.5%
Full CI	23m16s	14m10s	39.1%

CI feels responsive again#

Not every individual image build became faster—our free 2-vCPU build runners produced mixed timings—but the critical result is clear: the feedback loop developers experience is dramatically shorter.

Saving roughly nine minutes per run has made CI feel useful again instead of punitive. After spending far too long watching jobs sit in a queue, seeing the entire fan-out start in seconds is honestly delightful.

Data and methodology#

The measurements use GitHub Actions timestamps from September 1–3, 2026:

Full-CI medians: five successful main runs before migration and eleven after.
Pickup latency: 85 successful dependent jobs before and 110 after.
Job duration: completed_at - started_at.
Pickup latency: dependent job started_at - setup.completed_at.
Controlled baseline: GitHub-hosted run (https://github.com/kagent-dev/kagent/actions/runs/33634657352).
Controlled comparison: Blacksmith run (https://github.com/kagent-dev/kagent/actions/runs/33634940563).
Migration: PR #2659 (https://github.com/kagent-dev/kagent/pull/2659).

Thank you, Blacksmith#

We are grateful to Blacksmith for sponsoring kagent's CI. Their support helps us deliver faster, more reliable feedback to everyone contributing to the project.