InfrastructureAmazon Web Services2026-06-11AWS Bedrock Streams Tool Results in Real Time for Claude and Nova
Amazon Bedrock now supports real-time streaming of tool-use results for Claude 3.7 and Amazon Nova Pro, eliminating the batch-and-wait pattern that previously introduced latency into agentic loops. The feature is live today across all commercial Bedrock regions.
Original sourceAmazon Bedrock has added streaming tool-use support for Claude 3.7 Sonnet and Amazon Nova Pro, enabling model responses and tool call results to flow incrementally rather than waiting for a full round-trip to complete before proceeding. Previously, agentic workflows on Bedrock required the model to finish generating a tool call, the application to execute it, and the result to be batched back before the next generation step could begin — a pattern that compounded latency across multi-step tasks.
With streaming tool use, applications can begin processing partial tool results and surface intermediate state to users as the loop progresses. This is architecturally meaningful for any workflow with sequential tool calls: search-then-summarize, fetch-then-analyze, or multi-hop reasoning chains where each step depends on the previous. Sub-second agentic loops become feasible without custom buffering logic or side-channel workarounds.
The implementation follows the existing Bedrock Converse API surface, meaning applications already using streaming text generation can extend to tool use without significant refactoring. Support is scoped to Claude 3.7 and Nova Pro at launch, with no public roadmap for other models on the platform. The feature is available in all commercial AWS regions as of today with no separate opt-in required.
Panel Takes
The Builder
Developer Perspective
“The primitive here is clean: streaming tool results piped back through the same Converse API surface you're already using, no new SDK methods to learn. The DX bet was to not make this a separate API, and that was the right call — existing agentic code gets faster without a rewrite. The moment of truth is whether partial tool result handling is well-documented with real examples, not just a changelog entry and a handwave at the SDK reference.”
The Skeptic
Reality Check
“This is a real infrastructure gap that actually existed and was actually annoying, so credit where it's due — this isn't a fake feature. The narrow model support (Claude 3.7 and Nova Pro only at launch) is the thing to watch: if you're running Llama or Mistral on Bedrock, you're still batching, which means half the platform's users get the press release but not the feature. What kills this in 12 months isn't competition — it's AWS's own pace of expanding support to the rest of the model catalog.”
The Futurist
Big Picture
“The thesis this bets on is specific and falsifiable: that agentic workloads will move from experimental to production-latency-sensitive within 18 months, meaning infrastructure that was 'good enough' at 2-second loop times won't be. The second-order effect worth watching is that streaming tool use makes it practical to expose agentic loop state to end users in real time — which changes how products are designed around agents, not just how the backend performs. AWS is on-time to this trend, not early, but being on-time on infrastructure at AWS's distribution scale is still a meaningful forcing function for how other providers have to respond.”
The PM
Product Strategy
“The job-to-be-done is unambiguous: reduce perceived and actual latency in multi-step agentic workflows without forcing a migration to a different API. That's a single, complete job, and this feature nails it without adding configuration surface or new abstractions to learn. The gap worth flagging is model coverage — a PM building on Bedrock today has to architect around which models do and don't support streaming tool use, which reintroduces the complexity this feature was supposed to remove.”