Contained by default

Here's how the sandbox actually works.

No implicit trust, no shared access — just explicit boundaries for files, network access, tools, secrets, runtime limits, and artifacts, enforced on every single run.

Agent runtime
↓
Sandbox boundary
├─ filesystem rules
├─ network rules
├─ tool permissions
├─ secret access
└─ execution limits

Agents need room to work — not a blank check.

Production agents run code, read files, call tools, create artifacts, and touch real systems. Sandboxing gives them a controlled workspace — so the work gets done without exposing everything around it.

Unbounded execution
→ unclear file, network, and secret access
Sandboxed execution
→ explicit boundaries and traceable outputs

Define what the agent can reach.

Filesystem
The agent's workspace mounts read-only. Writes happen only through the sandbox boundary
Network
The agent container has no route out — every call passes through the sandbox
Tools
Exactly which tools the run is allowed to call
Secrets
Which credentials are injected and when
Runtime limits
Token usage, iteration count, and spend per run
Artifacts
What outputs can be retained, exported, or attached to traces

Every run gets its own workspace, with policy wrapped around it.

Every run carries a clear execution context: deployment version, runner, inputs, available tools, scoped knowledge, active skills, injected secrets, workspace boundaries.

Run workspace
├─ inputs
├─ temporary files
├─ tool outputs
├─ generated artifacts
└─ logs + trace events

Keep useful outputs without leaking the workspace.

Agents produce files, logs, summaries, reports, patches, and other outputs along the way. sypho separates what's retained from temporary execution state — and records artifact activity in the run trace.

run.started
workspace.created
tool.call: generate_report
artifact.created: report.pdf
artifact.attached
workspace.cleaned
run.completed

Boundaries that don't slow teams down.

Safer execution

File, network, tool, and secret access are scoped to each run.

Cleaner boundaries

Temporary workspace state stays separate from retained artifacts.

Environment control

Different sandbox rules apply across local, production, customer cloud, or on-prem.

Traceable outputs

Generated artifacts and workspace activity land in the run trace.

Production confidence

Agents work with real systems without unrestricted access.

Give agents room to work — with production boundaries.

Request access and help shape isolated execution for production agents.