Summary
agent.New() (Go SDK) hardcodes a 15-second timeout on the http.Client used for every outbound call the agent makes as a client — cross-agent srv.Call(...), and (separately) the control-plane memory backend's HTTP requests. There's no Config field to override it.
// sdk/go/agent/agent.go
httpClient := &http.Client{
Timeout: 15 * time.Second,
}Impact
Any reasoner chained behind srv.Call(...) that can legitimately take longer than 15 seconds end-to-end causes the caller to fail with a client-side timeout, even though the callee is still working and will complete successfully on its own. This is not theoretical: it reproduces immediately with any reasoning-model-backed reasoner (DeepSeek-R1/V3-style "thinking" models, o1-style models, etc.) doing a modest amount of work — a search call plus a large max_tokens reasoning response routinely exceeds 15s. The error surfaces as:
perform execute call: Post ".../api/v1/execute/<node>.<reasoner>": context deadline exceeded (Client.Timeout exceeded while awaiting headers)
This makes any multi-hop / recursive reasoner graph (agent calling agent calling agent) that uses a slower model unreliable by default, with no way to fix it from the caller's config.
Repro
- Register a serverless node with a reasoner that calls a reasoning-capable model with a generous
max_tokens budget (anything that reliably takes >15s wall-clock).
- Have a second reasoner invoke it via
srv.Call(ctx, "node.reasoner", input).
- The call fails with the context-deadline error above, while the callee's own execution (visible in the control plane) completes successfully after the caller has already given up.
Fix
Add a CallTimeout time.Duration field to agent.Config, defaulting to the existing 15s when unset (no behavior change for anyone not hitting this), and use it to construct the client's http.Client.
Acceptance criteria
Summary
agent.New()(Go SDK) hardcodes a 15-second timeout on thehttp.Clientused for every outbound call the agent makes as a client — cross-agentsrv.Call(...), and (separately) the control-plane memory backend's HTTP requests. There's noConfigfield to override it.Impact
Any reasoner chained behind
srv.Call(...)that can legitimately take longer than 15 seconds end-to-end causes the caller to fail with a client-side timeout, even though the callee is still working and will complete successfully on its own. This is not theoretical: it reproduces immediately with any reasoning-model-backed reasoner (DeepSeek-R1/V3-style "thinking" models, o1-style models, etc.) doing a modest amount of work — a search call plus a largemax_tokensreasoning response routinely exceeds 15s. The error surfaces as:This makes any multi-hop / recursive reasoner graph (agent calling agent calling agent) that uses a slower model unreliable by default, with no way to fix it from the caller's config.
Repro
max_tokensbudget (anything that reliably takes >15s wall-clock).srv.Call(ctx, "node.reasoner", input).Fix
Add a
CallTimeout time.Durationfield toagent.Config, defaulting to the existing 15s when unset (no behavior change for anyone not hitting this), and use it to construct the client'shttp.Client.Acceptance criteria
agent.Config.CallTimeoutexists; unset defaults to 15s (unchanged default behavior).http.Client.Timeoutused for outbound calls.