Self-hosted sandboxes and harness control planes that own the agent loop

Picrew

Implementation-first awesome list of agent harness engineering projects, tools, benchmarks, and guides with hundreds of GitHub entries.

Use caseUse it to discover obscure sandbox and harness layers mapped to execution, tooling, lifecycle, and governance needs.

cocoonstack

Fast cold-boot MicroVM sandboxes with warm pools, in-guest daemon, and Go/Python SDKs built on the cocoon stack.

Use caseUse it to claim sub-millisecond warm sandboxes or tens-of-ms golden clones for agent code execution.

Castor authors

Secure execution layer that treats every tool call as a syscall through a microkernel with budgets, immutable journal, pause, and resume.

Use caseUse it to enforce hard capability limits and budget-based approvals so agents cannot bypass structural gates.

Recurse team

Serverless agent harness plus coding-agent skill that iterates prompt and tool variants against verifiable domains with an FSM refinement loop.

Use caseUse it to auto-generate and evaluate specialist agents when you can define a clear input/output contract and eval.

Docker, Inc.

Docker shipped Cloud Sandboxes so agentic workloads continue in managed microVM isolation after the laptop sleeps. The same release advances Kits as OCI packages that bundle agent, tools, credentials, and network policy.

Shipped

Deanna Sparks

Docker's WeAreDevelopers keynote argues sandboxes plus external policy create the trust boundary agents need, and that Kits plus Cloud Sandboxes make authority reproducible and portable. The piece frames isolation and packaging as the next industrial layer for agentic work.

Read