Environment Pools

Environment Pools are now in public preview. Run coding and browser agent evaluations at scale without managing sandboxes, tunnels, or authenticated sessions.

Environment Pools

Four pool types ship today: Harbor (Docker sandboxes), Browser (cloud browsers with persistent profiles), OpenEnv (task apps), and Archipelago (production backends). Available to Pro and Team plan users.

python

from synth_ai.sdk.environment_pools import EnvironmentPoolsClient, PoolTask, AgentSpec
 
client = EnvironmentPoolsClient(api_key="sk_live_...")
 
# Create a pool
pool = await client.pools.create(
    name="my-pool",
    pool_type="harbor",
    capacity=10,
    concurrency=5,
)
 
# Add a task
task = PoolTask.from_docker(
    task_id="banking77",
    image="my-banking77:latest",
    command=["python", "task.py"],
)
await client.pools.tasks.create(pool_id=pool.id, task=task)
 
# Launch rollout
rollout = await client.rollouts.create(
    task_ref={"dataset": "banking77", "task_id": "banking77"},
    agent=AgentSpec.claude_code(model_id="claude-sonnet-4.5"),
    pool_id=pool.id,
    task_app_url="https://api.usesynth.ai/api/harbor/deployments/my-banking77",
    timeouts={"task_timeout_seconds": 300},
)
 
# Stream events with cursor resume
async for event in rollout.events():
    print(f"{event.event}: {event.data}")

Harbor Pools

Docker sandboxes for coding agents. Supports private registries with stored credentials.

python

# Store registry credential
await client.credentials.store(
    credential_type="docker_registry",
    name="ghcr",
    username="user",
    password="ghp_xxx",
)
 
# Use in pool
pool = await client.pools.create(
    name="private-pool",
    pool_type="harbor",
    registry_credential="ghcr",
)

Browser Pools

Pre-authenticated cloud browsers for web automation. Persistent profiles, stealth mode, CDP access, session reuse.

python

task = PoolTask.from_browser(
    task_id="linkedin-scrape",
    task_app_url="https://my-task-app.com",
    profile="linkedin",  # stays logged in across rollouts
)

Rollout Lifecycle

RolloutHandle provides status, events, artifacts, usage, and cancellation:

python

rollout = await client.rollouts.create(...)
 
# Wait for completion
await rollout.wait()
 
# Check results
status = await rollout.get()
print(status.phase)  # succeeded, failed, canceled
 
# Download artifacts
await rollout.download_zip("artifacts.zip")
 
# View usage
usage = await rollout.usage()
print(f"Cost: ${usage.estimated_cost}")
print(f"Sandbox minutes: {usage.sandbox_minutes}")

SSE Streaming

Cursor-based event streaming with auto-reconnect:

python

async for event in rollout.events(cursor="last-event-id"):
    if event.event == "lifecycle":
        print(f"Phase: {event.data.phase}")
    elif event.event == "reward":
        print(f"Reward: {event.data.value}")

Capabilities

Query supported agents, models, and backends:

python

caps = await client.capabilities.get()
print(caps.agents)   # ["claude-code", "opencode", "codex"]
print(caps.models)   # ["claude-sonnet-4.5", "gpt-5.2-codex", ...]
print(caps.backends) # ["harbor", "browser", "openenv", "archipelago"]

Usage Tracking

Per-rollout metering for sandbox minutes, browser minutes, LLM tokens, and data transfer. View usage via rollout.usage().

Other Fixes

Plan gating now shows upgrade links in error messages
Artifact download performance improvements
Browser pool health monitoring
SSE reconnection with exponential backoff

Environment Pools

python

from synth_ai.sdk.environment_pools import EnvironmentPoolsClient, PoolTask, AgentSpec
 
client = EnvironmentPoolsClient(api_key="sk_live_...")
 
# Create a pool
pool = await client.pools.create(
    name="my-pool",
    pool_type="harbor",
    capacity=10,
    concurrency=5,
)
 
# Add a task
task = PoolTask.from_docker(
    task_id="banking77",
    image="my-banking77:latest",
    command=["python", "task.py"],
)
await client.pools.tasks.create(pool_id=pool.id, task=task)
 
# Launch rollout
rollout = await client.rollouts.create(
    task_ref={"dataset": "banking77", "task_id": "banking77"},
    agent=AgentSpec.claude_code(model_id="claude-sonnet-4.5"),
    pool_id=pool.id,
    task_app_url="https://api.usesynth.ai/api/harbor/deployments/my-banking77",
    timeouts={"task_timeout_seconds": 300},
)
 
# Stream events with cursor resume
async for event in rollout.events():
    print(f"{event.event}: {event.data}")

Harbor Pools

Docker sandboxes for coding agents. Supports private registries with stored credentials.

python

# Store registry credential
await client.credentials.store(
    credential_type="docker_registry",
    name="ghcr",
    username="user",
    password="ghp_xxx",
)
 
# Use in pool
pool = await client.pools.create(
    name="private-pool",
    pool_type="harbor",
    registry_credential="ghcr",
)

Rollout Lifecycle

RolloutHandle provides status, events, artifacts, usage, and cancellation:

python

rollout = await client.rollouts.create(...)
 
# Wait for completion
await rollout.wait()
 
# Check results
status = await rollout.get()
print(status.phase)  # succeeded, failed, canceled
 
# Download artifacts
await rollout.download_zip("artifacts.zip")
 
# View usage
usage = await rollout.usage()
print(f"Cost: ${usage.estimated_cost}")
print(f"Sandbox minutes: {usage.sandbox_minutes}")

async for event in rollout.events(cursor="last-event-id"): if event.event == "lifecycle": print(f"Phase: {event.data.phase}") elif event.event == "reward": print(f"Reward: {event.data.value}")

caps = await client.capabilities.get() print(caps.agents) # ["claude-code", "opencode", "codex"] print(caps.models) # ["claude-sonnet-4.5", "gpt-5.2-codex", ...] print(caps.backends) # ["harbor", "browser", "openenv", "archipelago"]