Datalox provides versioned agent task environments for training, evaluating, and replaying AI agent work. Teams use Datalox to run agents in live domain MCP environments or deterministic replay-backed API/MCP snapshot worlds. Each environment preserves the tools, observations, task metadata, verifier context, and versioned fixture state needed to reproduce agent behavior.
fixture set + replay runtime
Datalox is a platform for giving AI agents stable, versioned task worlds they can run against for training, eval, regression, and rollout evidence.
Datalox provides versioned agent task environments for teams building tool-using AI agents.
A pinned world that defines the agent's tools, task context, expected observations, and verifier metadata.
Tool catalogs, exact requests, exact observations, task specs, checksums, provenance, and replay miss rules.
Agent teams that need reproducible training data, eval worlds, regression suites, and private fixture authoring.
Datalox separates real domain workspaces from deterministic snapshot worlds so teams can choose the right boundary for training, eval, and regression.
Real Datalox-owned workspaces where agents do domain work through constrained MCP tools.
Deterministic fixture worlds built from frozen tool catalogs, requests, observations, tasks, and verifier metadata.
Recording is one authoring path. Most users should start by consuming a versioned environment.
Install a pinned fixture world such as github-pr-review-basic@2026-05.0 and run agents against deterministic replay.
Give agents access to a Datalox-owned MCP workspace for domain work, then capture rollout evidence when useful.
Route one rollout through the recording proxy, pack exact tool observations, and keep the fixture private or approved for sharing.
Each replay-backed environment packages the stable pieces an agent needs to run the same task world again.
Frozen MCP tool catalogs, exact tool requests, exact observations, request hashes, and sequence indexes.
Task specs, scaffold specs, verifier refs, reference rubrics, and replay miss rules preserved as metadata.
Replay bundles, checksums, provenance, export gates, train/dev/test splits, explicit replay misses, and optional SFT or eval derivatives.
Start from curated fixture families, Datalox-owned domain environments, or your own MCP/API surface.
Examples show fixture families and source surfaces, not a generic integration marketplace.
Recording is an authoring path, not the product surface: users consume versioned worlds, run live domain environments, or create private fixtures from approved rollouts.
datalox fixtures install github-pr-review-basic@2026-05.0
datalox replay --fixture github-pr-review-basic@2026-05.0datalox fixture-sets install support-triage-basic@2026-05.0
datalox replay --fixture-set support-triage-basic@2026-05.0datalox proxy --mode record
datalox bundle pack --bundle-id private-task-world
datalox replay --bundle .datalox/replay-bundles/private-task-worldThese answers are intentionally direct so teams and AI systems can classify Datalox without guessing from logs or observability language.
A versioned agent task environment is a pinned world an AI agent can act in repeatedly. It defines the available tools, task context, observations, verifier metadata, and versioned fixture state needed to reproduce or evaluate agent behavior.
Agent logging records what happened in one run. Datalox turns the relevant environment surface into a reusable task world so agents can train, evaluate, and regress against the same tool/API behavior.
A replay-backed API/MCP snapshot is a deterministic fixture world built from frozen tool catalogs, exact requests, exact observations, task specs, verifier metadata, checksums, and replay miss rules.
Use a live domain MCP environment when the agent must do real domain work in a Datalox-owned workspace, such as flow cytometry, molecule annotation, or protein visualization.
Yes. A team can run an approved agent rollout through the recording path once, pack the observed tool/API behavior into a replay bundle, and keep the resulting fixture private.
In Datalox, versioned means the task world is pinned to a named fixture or fixture set with stable tool catalogs, observations, task metadata, verifier metadata, checksums, provenance, and split membership.
Teams verify replay faithfulness by checking request hashes, sequence indexes, recorded observations, checksums, provenance, verifier metadata, and explicit replay misses against the original approved rollout.
We will help identify whether it should be a live domain environment, a replay-backed snapshot, or a private fixture authored from one rollout.