A browser and research agent built for the part demos usually skip:
preserving progress, reducing noisy pages into useful observations, and
checking whether the requested result actually happened.
3,055tests passed
8a6a8c7public commit verified
27 secreal-web demo
openMIT source
cinematic real-web proof · one continuous browser storyexact public commit · independently checked outcome
What this proves
The recording shows Blackreach's real public browser code operating a
real website. Animated overlays are synchronized to the captured run so
the goal, observation, action, and final verification remain readable
without a split screen. The decisions are deliberately scripted, so this
is product proof, not an autonomous model evaluation.
The reliability gap
Browser-agent demos often assume a clean page, a short task, and a perfect
network. Real research work is messier. Pages render late, sessions expire,
layouts change, and a technically successful request can still contain the
wrong result.
I built Blackreach to keep that uncertainty inside the system. An action
trace is useful evidence, but it is not enough. Progress has to survive
interruption, and the final state needs its own check.
Three engineering decisions
Reduce the observation
Turn a complex live page into the smaller set of facts and controls needed for the next decision.
Preserve honest state
Keep interrupted work inspectable and resumable without relabeling a partial run as complete.
Check the outcome
Judge the resulting page or artifact separately from the agent's own completion report.
Expose uncertainty
Treat repeated actions, unexpected content, and incomplete results as visible failures to investigate.
Evidence, not just claims
I reran the exact public GitHub commit shown in the video on July 26,
2026. Its full test suite completed with 3,055 passing tests. The
controlled run then imported Blackreach from that same checkout, searched
Wikipedia in a real browser, followed an observed link, and verified the
visible destination heading.
SourcePublic commit 8a6a8c7, imported directly for the run.
Suite3,055 tests passed in a fresh run against that public commit.
DemoReal browser interaction on Wikipedia, with the resulting page heading checked separately.
LimitThis is a deterministic product proof, not a benchmark of model planning quality or performance on the open web.
What stays private
The public repository and this case study show the product surface,
testing discipline, and engineering judgment. They do not publish private
deployment details, browsing environments, session material, operational
weaknesses, or a reproducible playbook for the private system.
The lesson reaches beyond browser agents: autonomous tools earn trust by
preserving state, checking outcomes, and making weak evidence visible.
Need someone who can make agent systems survive real runs?I am open to engineering roles and focused contract work.