Run the version you chose
A saved harness, its prepared runtime, and the selected benchmark inputs define the experiment. Their identities remain attached to the run, so later edits do not silently change what was tested.
Trialground Runtime
Trialground
Disposable sandbox computers for agent experimentation.
Run agent harnesses with the tools and worlds they are configured to use. Trialground prepares dependencies, manages execution across benchmark cases, and retains outputs and recorded trajectories so your team can investigate behavior and compare candidates.
Trialground Runtime
Keep the experiment connected to what you ran.
An agent’s behavior depends on its harness, inputs, tools, and environment. Trialground keeps the selected candidate and benchmark inputs attached to execution, giving your team a stable basis for investigating the results.
Reuse prepared dependencies
Trialground prepares an execution image from your harness’s declared Python dependencies and reuses it for compatible runs. Your team can iterate on the harness while the platform manages environment preparation and execution.
Continue through interrupted work
Accepted work and completed results have durable records. Trialground reconciles execution progress and recovers eligible interrupted work, keeping completed batches available as the rest of the experiment proceeds.
Keep evidence after execution
Outputs, execution status, and recorded trajectories remain available after the sandbox finishes. Infrastructure failures are distinguished from harness failures, so missing execution evidence does not become a misleading judgment of the candidate.
Agent experiments
Put your agents through the work they need to handle.
Start with a harness and the behavior you need to examine. Trialground supports development work across benchmark conditions, candidate implementations, and repeated trials.
Exercise the conditions your agent must handle
Run a harness across the tasks, inputs, and exceptions represented in your benchmark. Repeated trials help examine behavior that a single successful run cannot establish.
Compare changes to the harness
Test candidate implementations against selected benchmark inputs. Keep the implementation and execution context behind each result available as you investigate differences.
Test agents working with tools and worlds
Exercise the tools and worlds your harness is configured to use, and inspect the recorded interactions. World state, tool access, and reset behavior are defined by that setup; the disposable computer supplies the harness runtime.
Support an agent’s optimization work
Lemon Goal uses Trialground to test competing harness prototypes against a benchmark objective. Execution records give each candidate investigation a concrete result to build on.
Why state, tool behavior, and the full trajectory matter when assessing an operational agent.
Runtime comparison
Choosing infrastructure for agent experiments.
Compare Trialground with sandboxes for agents like Daytona, E2B and Modal. Consider it when your workload involves repeated trials across tasks, world conditions, and candidate implementations, with execution organized around your benchmarks.
These platforms also support agent testing. Trialground’s focus is the managed experiment: executing supported harnesses against selected cases, preserving the candidate identity, and retaining the resulting evidence. Compare the execution work your team would need to assemble for the workload you have.
| Decision | Trialground Runtime | Across your shortlist |
|---|---|---|
| How experiments are defined | Saved harness versions and pinned benchmark cases form the basis of managed execution. | How your chosen runtime and evaluation framework connect candidates, inputs, and repetitions. |
| Who manages execution | The platform prepares dependencies, admits execution jobs, and reconciles their progress and results. | The setup and orchestration your team needs, along with the required runtime flexibility and capacity. |
| What remains after a run | Bounded outputs, execution status, and captured trajectories stay connected to the experiment. | The evidence you need to collect, retain, and associate with each candidate and task. |
Choose around the runtime flexibility and scale you need. For production serving, broader application hosting, or requirements beyond Trialground’s supported harness profiles, assess those needs separately. Review the runtime requirements ↓
Trialground Evaluation
Establish which standards the behavior meets.
Apply expert rubrics to outputs and recorded trajectories. Choose from supported judge models, including Kestel, our specialized binary rubric judge. Compare candidates by policy, rubric, and coverage, with each judgment connected to its criteria and evidence.
A completed agent task can still rely on insufficient evidence, take an unacceptable action, or miss a requirement. Rubric evaluation makes those distinctions available for review.
Trialground EvaluationInteractive product demo
Loading Benchmark Eval…
Compare candidate behavior by case, policy, and rubric.
Trialground EvaluationJudgments, criteria, and recorded behavior
Rubric evaluation
Assess the evidence against the applicable requirement.
Response rubrics assess the final response; trajectory rubrics assess behavior visible in recorded execution. Review the criterion alongside the behavior and its judgment. This makes it possible to investigate why a completed task failed the expert standard.
- Response
- What the final answer demonstrates
- Trajectory
- What the recorded execution demonstrates
- Judgment
- The applicable criterion and the supporting evidence
“Some evaluation tasks require specialized domain expertise. If the AI’s judgments don’t agree with those experts, we may not want to rely on it for that task.”
Execution and assessment stay connected.
Trialground Runtime produces the execution record. Trialground Evaluation applies your criteria to the captured behavior. Keep the harness, cases, and assessment basis in view when deciding whether a candidate improved.
Runtime requirements
Match the environment to the experiment.
Bring the harness, inputs, dependencies, and tool or world requirements involved. Establish the execution conditions before using the results to compare candidates.
Harnesses and dependencies
The current managed runtime supports saved Python 3.12 harnesses on Linux, with declared Python dependencies and JSON-compatible results. The harness defines how the agent uses its models, tools, and task context.
Execution and state
A disposable execution job can process several benchmark inputs in the same Python process and workspace. State isolation between cases and reset behavior for connected worlds need to be addressed in the harness setup.
Capacity and limits
Concurrent execution is admitted within workspace and platform limits. Compute profiles, case counts, repetitions, execution time, and output limits determine the workload to plan. Confirm the available capacity for your program before scheduling it.
Recorded evidence
Trajectory capture depends on the supported instrumentation and its configured limits. Outputs and recorded events are retained separately from the execution environment. Files created inside a sandbox need an explicit persistence path if they must remain available.
Runtime and rubric evaluation have separate usage meters. The subscribed scope and agreed profiles define what is available for your program. Review pricing →
Within Teammately
Put the evidence to work.
Use trial results to direct a harness change, refine the benchmark, or clarify the standard. Runtime and Evaluation connect to the same development foundation.
Start with your workload
Bring the agent you want to test.
Start with a saved harness and the inputs you want to exercise. We can work through the dependencies, tools, worlds, and runtime requirements, then establish the comparisons or rubric assessments your development question needs.