🚀 Announcing Agent Sandbox v1.0.2!

We're excited to announce the release of Agent Sandbox v1.0.2! This release brings significant improvements across the sandbox router, SDKs, and reinforcement learning (RL) integrations. Key enhancements include Ed25519-based scoped-token v2 verification for fine-grained authorization, direct in-cluster connectivity options for the Go SDK, improved connection resiliency for the Python SDK, warm pool adoption and OpenHands integration in the RL framework, and controller observability metrics via controller-runtime v0.25.0.

⚠️ Breaking Changes / Action Required

Sandbox Router Authorizer Interface Update (authz.Authorizer) (#1497): The public Authorizer interface signature in Go sandbox-router has changed from taking separate (namespace, name) string arguments to accepting a single authz.AuthorizationTarget struct. Custom Authorizer implementations must update their Authorize method to inspect the fields of authz.AuthorizationTarget.
Sandbox Router Namespace Change (#1537): The Go sandbox-router manifests in sandbox-router/deploy/ have moved from the default namespace to agent-sandbox-system to enforce privilege separation and align with SDK defaults and NetworkPolicy rules. If you deploy raw manifests, ensure your manifests and scripts point to agent-sandbox-system.

Key Highlights

Sandbox Router & Security

Scoped-Token v2 Target Binding (#1497): Added Ed25519 scoped-token v2 verification binding tokens to Sandbox UID, port, HTTP method, and exact upstream path. Enables multi-key rotation and exclusive cutoff timestamps (--authz-scoped-token-v1-accept-until) for seamless migration from legacy HMAC v1 tokens.
Namespace & NetworkPolicy Alignment (#1537, #1539): Deployed the router to agent-sandbox-system and added the app: sandbox-router label to router deployment manifests, ensuring compatibility with the controller's secure-by-default SandboxTemplate NetworkPolicy.
OLM Operator Router Integration (#1425, #1477): Added the sandbox-router deployment, proxy service, and health checks directly to the Operator Lifecycle Manager (OLM) bundle.

Go & Python SDKs

Direct In-Cluster Connectivity in Go SDK (#1576): Added Options.Connectivity (ConnectivityInClusterService and ConnectivityInClusterPodIP), enabling workloads inside the cluster to connect directly to sandbox pods via headless service DNS or Pod IP without proxying through the API server or router.
Python SDK Connection Resilience on 4xx (#1574): Fixed an issue where HTTP 4xx responses (such as file not found) caused the connector to tear down the port-forward tunnel and clear the cached Pod IP. Tunnel teardown and routing invalidations are now reserved for transport failures and 5xx errors.
Async Client Concurrency & Flake Hardening (#1385, #1553): Added comprehensive E2E tests for AsyncSandboxClient validating concurrent execution and non-blocking event loops, with interval-overlap assertions resilient to sub-second warm-pool creation speeds.

Reinforcement Learning (RL) & Fleet Integrations

Warm Pool Adoption (#1556): Added FleetConfig.adopt_existing=True and FleetConfig.pool_name_format to the RL SDK, allowing training harnesses to discover and adopt pre-provisioned warm pools by container image rather than creating redundant pools.
OpenHands Fleet Adapter & Fan-Out (#1503): Introduced OpenHands integration shims (make_fleet_workspace, make_handle_workspace), advisory attach-time version skew detection, and scalable multi-agent example workflows (examples/run_openhands_fleet.py).
RL Quickstart Standardization (#1549): Updated setup steps to configure and provision workloads consistently in the agent-sandbox-rl namespace.

Controller Observability, Performance & MCP Server

controller-runtime v0.25.0 & REST Client Metrics (#1547): Upgraded controller-runtime to v0.25.0 and enabled client-go REST client metrics (rest_client_requests_total, rest_client_request_duration_seconds, rate-limiting, and retry counters) by default.
MCP Server Probes & Skills (#1339, #1534): Added /healthz (liveness) and /readyz (readiness) probe endpoints to the Model Context Protocol (MCP) server container and documented client skills for MCP tools.
Golden-Snapshot Warm Pools on GKE (#1522): Added an end-to-end example demonstrating golden-image warm pools restored directly from GKE Pod Snapshots.
High-Throughput Tuning & Diagnostics (#1593, #1552, #1555): Documented flags for connection sharding, write optimization, and refill rate limiting; tuned stress test networking defaults (nodeCIDRMaskSize=23); and added serial port log capture during benchmark cluster validation failures.

Installation

Standard Install (Core + Extensions)

Recommended for most users and GitOps engines (Argo CD, Config Sync, kustomize):

kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/download/v1.0.2/sandbox-with-extensions.yaml

Selective Install

Install components separately:

# Core only:
kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/download/v1.0.2/sandbox.yaml

# Extensions (opt-in):
kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/download/v1.0.2/extensions.yaml

Python SDK

pip install k8s-agent-sandbox==1.0.2

Contributors

We extend our sincere thanks to all contributors to this release:
@aditya-shantanu, @dongjiang1989, @drogovozDP, @esposem, @hyperb1iss, @janetkuo, @justinsb, @leomcl, @nankeen, @tomergee, @vicentefb, @vincent0426, @wjun29, @yuzhiquan

New Contributors

@hyperb1iss made their first contribution in #1497
@leomcl made their first contribution in #1463
@nankeen made their first contribution in #1576

Full Changelog: v1.0.1...v1.0.2

kubernetes-sigs/agent-sandbox v1.0.2 on GitHub