← Back to feed

OpenRouter’s May 11 changelog ships two features that address the gap between demo-grade agents and production deployments: human-in-the-loop (HITL) tool calling and session-id-based provider stickiness. Before these additions, teams deploying agents in sensitive domains were building both layers themselves, or working around them by calling provider APIs directly.

Human-in-the-Loop Tool Calling

The HITL system hooks into OpenRouter’s Agent SDK at the tool-call level. Tools define an onToolCalled handler. When the handler returns null, the agent loop halts with status: 'awaiting_hitl' and surfaces the pending call to whatever review surface the developer has wired up — a UI, CLI queue, API endpoint, or messaging thread. The developer collects a human decision and resumes by calling callModel again with a function_call_output item for each paused call.

Two modes are available:

Conditional escalation. The onToolCalled handler receives parsed inputs and decides per-call whether to auto-resolve or pause. A deployment tool could auto-approve staging pushes and escalate production pushes. A financial agent could auto-approve reads and pause on writes above a dollar threshold. The logic lives in one place and the resume path is identical regardless of what triggered the pause.

Pre-execution approval. The requireApproval flag gates tool execution entirely before the tool logic runs, rather than intercepting after. Use it when the consent gate matters more than the tool’s output — approving whether a tool should execute rather than reviewing what it returned.

An onResponseReceived hook fires on human-supplied output before the model sees it, allowing enrichment or validation: adding audit timestamps, clamping values to valid ranges, or translating freeform human responses into the tool’s expected output schema.

The system is model-agnostic. Any model available through OpenRouter works with HITL tools.

Session-ID Provider Stickiness

Multi-turn agentic workflows have an underappreciated cache problem. When consecutive requests in a conversation route to different provider instances, KV cache misses increase — the model re-encodes prefix tokens on every turn instead of reusing cached activations. For long agent sessions with substantial shared context, this adds latency and cost at every turn.

The session stickiness fix routes requests sharing the same session-id header to the same provider when possible. OpenRouter describes the feature as best-effort: availability constraints and load balancing can still route a session elsewhere, but stickiness is now a first-class optimisation signal. Cache hit rates improve for the multi-turn agentic workflows that now account for a disproportionate share of frontier API spend.

Other May 11 Changes

Workspace URLs now resolve via immutable workspace IDs, so shared links survive slug renames. Model page headers were rebuilt for faster scanning. Provider cards on model pages surface flex and priority pricing tier differences as compact pills at a glance. A bug in cost_details that miscounted costs when server tools ran multiple inner turns is fixed.

Why It Matters

The HITL addition is the architecturally significant one. Every production agent deployment that touches sensitive operations — finance, legal, healthcare, infrastructure management — needs human approval gates on high-stakes tool calls. Before this, teams were implementing that layer themselves or adopting opinionated orchestration frameworks like LangGraph that bake in their own state machines. OpenRouter’s approach keeps the primitive thin: pause and resume at the tool level, no mandatory framework adoption.

Session stickiness is table stakes for enterprise deployments. Its absence was a reason to avoid OpenRouter for long-running agent workflows and go directly to provider APIs where session management is more predictable. The fix closes that gap.

Together the two features push OpenRouter toward production agent infrastructure rather than an API routing layer. The distinction matters for the competitive dynamic: Anthropic, OpenAI, and Google are all building their own managed agent products with native session handling and approval mechanisms. OpenRouter’s bet is that a model-agnostic, provider-agnostic layer below those products is where enterprise agent deployments land at scale.