Sandboxing / controlled machines for agent evaluation
Give agents room to work. Keep the boundary.
Run terminal and code-execution workloads inside isolated, image-backed environments—with a complete tool surface and evidence for every execution.
terminal-agent / ubuntu:24.04
IMMUTABLE IMAGE · STANDARD-4
filesystem
/app mounted and ready
command_execute
python /app/evaluate.py
network
environment policy applied
lease
released when evaluation completes
08
COMMANDS
03
ARTIFACTS
00
VIOLATIONS
RUNTIME
Isolated
container-backed execution boundary
TOOL SURFACE
Files + shell
commands, processes, and terminals
PUBLIC DEMO
Coming soon
guided sandbox and technical write-up
01 / Real environment isolation
Give every workload its own controlled runtime.
Run terminal and code-execution evaluations inside isolated container environments instead of handing an agent access to your machine. Each environment has an explicit image, compute profile, network policy, and scoped lease.
- Public container images resolved into reproducible deployments
- Explicit compute, storage, network, and lifecycle policy
- Environment access scoped to an authorized execution lease
Runtime image
resolved to immutable source digest
Compute profile
explicit CPU, memory, and storage
Network policy
declared for this environment
Execution lease
scoped capability and expiration
02 / A complete execution surface
Let agents work through tools you can observe.
Expose the machine as a structured capability surface. Models can inspect files, execute commands, manage processes, and use an interactive terminal while Trunchbull records the operations required to evaluate the result.
- File read, write, list, move, and delete operations
- Command execution and managed background processes
- Interactive terminal sessions with controlled input and output
file_read
Inspect workspace content
file_write
Create evaluation artifacts
command_execute
Run bounded shell commands
process_list
Inspect managed processes
terminal_open
Start an interactive session
03 / Public demo space
Try the sandbox yourself soon.
We are preparing a public space where you can inspect the environment, run a constrained task, and see the same execution evidence Trunchbull captures during a benchmark.
- No local installation required
- Guided terminal and filesystem task
- Technical architecture write-up to follow
GUIDED ENVIRONMENT / 01
Open a terminal. Complete a task. Inspect the boundary.
A public, constrained sandbox with the execution trace visible beside it.
From image to evidence
A controlled environment in three moves.
Choose the runtime, grant the execution a scoped lease, and inspect the recorded work before the environment is released.
- 01
Define the environment
Select an image and explicit compute, storage, and network policy.
- 02
Lease scoped access
Give the evaluation only the sandbox capabilities it requires.
- 03
Execute and verify
Record operations, inspect artifacts, grade the result, and release the lease.
Public demo coming soon
See the boundary, not just the benchmark.
The public sandbox and a detailed architecture write-up are next. Until then, create a workspace to run controlled environments with your own evaluations.