ReplayAutonomous app testing for teams who ship at AI-speed
Promoted
Warren runs agent harnesses as isolated, observable workloads on infrastructure you control. It owns the workspace, limits, live events, recovery, and Git delivery.
What's great
the "operator can cancel any live run, supported runtimes also accept mid-run steering" part is what makes this feel like real infrastructure instead of a wrapper script. most agent runners I've tried treat a run as fire-and-forget until it finishes or dies - having per-run and per-project spend caps plus the ability to reconcile a failed runtime and salvage the work already done is exactly what's missing once you're running more than one agent at a time.
What needs improvement
getting from zero to a first run assumes you're comfortable choosing between local, a sibling Docker container, or a k8s pod up front. a bare "just run it on my laptop, no Docker" path would lower the bar for someone who just wants to see the run history and event stream before deciding how to actually deploy it.
vs Alternatives
looked at just running agents in tmux on a rented box, and separately at Kubernetes jobs directly. tmux gives you none of the run history or spend tracking, and raw k8s jobs don't give you the Git delivery step or mid-run steering without building that yourself. Warren is the middle layer that was missing.
Ratings
Ease of use
Reliability
Value for money
Customization
Report
6 views
Warren
Went and looked at the public run history and the useful part is the cost column. On your own warren repo the runs from the last eight hours go from $0.538 to $4.47, same project, same day. That spread is the thing nobody building on top of agents can price around, and publishing it with no login does more for me than the seven bullets above it.
Warren
@asadmalik901 Glad you find it to be useful! That spread is why the cost column is public. Agent runs are variable workloads: scope, model, retries, and tool use can produce materially different costs, even with the same model in the same repo on the same day.
Warren records actual spend and supports per-run and project-level caps, so operators can plan around observed distributions rather than a misleading average.
@jayminwest The caps are the part I'd actually use. My expensive runs aren't the $4.47 ones, they're the retry loops that burn twenty minutes before anyone notices something's stuck. A project cap that kills the run instead of emailing me about it is worth more than the whole dashboard.