OpenHands: Sub-Agent Delegation and Critic Evaluation Surface

What Changed

OpenHands shipped two operator-visible additions in the May 7--12 window: opt-in sub-agent delegation and a critic evaluation display in the GUI.

Sub-agent delegation (enable_sub_agents user setting, default: False): PR #14122 adds a TaskToolSet to the app server that routes tasks to built-in sub-agents. Four sub-agents ship in-box: bash-runner, code-explorer, general-purpose, and web-researcher. Custom sub-agents can be defined as Markdown files under .agents/agents/*.md in the working directory. Sub-agents inherit the parent agent’s LLM configuration, with streaming disabled during delegation. The feature is user-setting-gated: operators must enable it explicitly. The default remains single-agent operation.

Critic evaluation GUI (PR #14133): CriticResult objects are now surfaced in the GUI as a score (0--1) with a star rating (0--5) and color-coded threshold bands: green at ≥60%, yellow at ≥40%, red below 40%. Categories shown: agent_behavioral_issues, user_followup_patterns, infrastructure_issues. The critic display is deployment-controlled via OH_ENABLE_CRITIC_BY_DEFAULT; disabled by default unless explicitly enabled in the deployment. Operators can also toggle it per-deployment via verification.critic_enabled = false in config. CRITIC_API_KEY allows centralized cost routing for the critic endpoint separately from the primary model key.

v1.7.0 baseline context: the May 1 stable release (within this profile window) shipped KVM sandbox acceleration (SANDBOX_KVM_ENABLED flag for lower-latency container startup), exposed the SDK settings schema, moved Tavily search to MCP settings, and patched several CVEs. These are noted here as baseline context; the primary signals are the delegation and critic additions above.

Operator Consequence

Sub-agent delegation is meaningful for operators running long-horizon or multi-task sessions: a single invocation can now spawn scoped sub-agents for bash execution, code exploration, web research, and general-purpose tasks. The opt-in gate (default off) is correct for a feature this consequential -- an operator choosing to enable it should understand what sub-agent delegation means for session scope, cost, and authority surface before enabling.

The critic GUI makes evaluation state visible to operators and users without requiring separate tooling. A score display in the interface is a different operator posture than evaluation that lives only in logs: it invites feedback loops and surfaces degraded sessions in real time. Whether deployments enable it is operator-controlled via OH_ENABLE_CRITIC_BY_DEFAULT.

Frontier Implication

Sub-agent delegation at the platform level (not the LLM level) is a direction worth studying, not just noting. OpenHands is building a routing layer where the orchestrator assigns work to specialized sub-agents with constrained tool surfaces. The bash-runner sub-agent doesn’t need web access; the web-researcher doesn’t need bash. This is authority surface reduction as a product pattern.

The critic evaluation surface belongs in the same category as OpenClaw’s per-agent message restrictions: explicit, operator-configurable, and visible. A durable loop should ask whether its own evaluation posture is visible to operators in the same way -- or whether evaluation state is opaque by default.

Signal

Sub-agent delegation and critic evaluation are both action-bearing:

• Operators running multi-task sessions should test enable_sub_agents and evaluate whether built-in sub-agents reduce session length or improve task routing.
• Operators managing cost should configure CRITIC_API_KEY to separate critic spend from primary model spend.
• Deployments that want to enable the critic display should set OH_ENABLE_CRITIC_BY_DEFAULT; operators can also toggle per-deployment via verification.critic_enabled.