A voice agent can sound fluent during a short demo and still fail a normal support call. The customer corrects a serial number, a diagnostic tool takes several minutes, the audio connection renews, and the agent returns to a question the customer already answered. The expensive defect is the lost thread of the interaction.
Amazon’s October 5, 2026 Nova 2.5 Sonic release and Strands Bidi general availability deserve a joint assessment because they address two layers of this problem: the speech model’s handling of work and the framework’s handling of an ongoing conversation. Both are new coverage in the Amazon research series; the release date precedes our earlier Bedrock Managed Agents assessment, which covered a separate AWS/OpenAI preview.
My thesis is that the adoption experiment should center on continuity across connection renewal. That is where conversational state, unfinished tools, elapsed time, and user intent must remain consistent.
What the October release establishes
AWS announced Nova 2.5 Sonic as generally available in four Bedrock regions, with improved reasoning, instruction following, tool calling, and latency at the same pricing as Nova 2 Sonic. Those improvements are vendor claims, without an independent measurement in this article. The launch retains seven-language speech, mixed voice/text sessions, and asynchronous tools. Nova 2.5 Sonic announcement
Strands Bidi’s GA announcement describes proactive provider reconnection, carrying context through the provider’s supported mechanism and seeking a turn boundary before a bounded deadline. It names an eight-minute Nova connection limit. It also documents OpenTelemetry events, tool integration, interruption handling, echo cancellation, and provider adapters. Persistence across an application restart is explicitly a separate concern. Bidi GA announcement
The distinction matters: a longer conversation can span several connections. That does not make every underlying resource indefinite, nor prove that every fact and pending operation survives identically.
A context-limit discrepancy remains unresolved
The dated AWS launch says 256K context. The current Nova models page says up to 1M tokens for Nova 2.5 Sonic. Both statements were visible when checked on October 19. Neither page, as inspected, reconciles that difference. Dated launch, living model page
We should preserve that discrepancy instead of silently adopting the larger number. Before capacity planning, obtain the applicable limit for the selected model identifier, region, API, and modality accounting. Then verify the actual workload’s behavior near that limit. A connection-duration limit and a token-context limit are different constraints; relaxing either one does not automatically relax the other.
This is also a useful evaluation discipline. A long audio session should be described by elapsed duration, input and output usage, retained facts, tool events, and number of reconnections. “Long context” alone tells a reader too little about what happened.
Native speech changes what the transcript can prove
AWS’s service card describes speech encoding and rendering around a multimodal language model, with pretraining, supervised fine-tuning, and reinforcement learning. It says the successor keeps the prior API surface, voices, and languages while improving instruction following and tools, and recommends retesting each use case. Nova 2.5 Sonic service card
An application evaluation should therefore retain audio as an experimental input, with appropriate permission and retention controls. A text transcript is useful for inspection, but cannot represent every pause, interruption, pronunciation, or overlapping turn. Replaying only text changes the experiment.
For an illustrative diagnostic assistant, spoken “fifteen” and “fifty” can produce different warranty lookups. A transcript might reveal the final recognized value while hiding the hesitation that led the model to ask a clarifying question. Measure whether the assistant resolves the uncertainty correctly, alongside the readability of the transcript.
Avoid using perceived stress as authenticated identity or as evidence that a customer’s request should override policy. Emotional responsiveness and transaction correctness deserve separate labels in the evaluation. A reassuring voice can make a wrong answer feel more authoritative.
Reconnection is a state-consistency experiment
Consider a proposed appliance-support call. The customer supplies a product serial number, rules out one troubleshooting step, and asks the agent to check replacement eligibility. That lookup continues while the customer discusses delivery options. The connection renews before the result arrives.
The evaluator should establish whether the new connection retains the corrected serial number, the rejected troubleshooting step, the pending lookup’s identity, and the customer’s latest preference. An agent that restarts politely but loses those facts has preserved audio service without preserving the task.
Keep an application record of confirmed entities and tool operations that can be compared with the conversation. This record is an evaluation aid and a source of operational truth; it should not silently transform an uncertain utterance into a confirmed value. Record corrections explicitly.
A result from an older lookup can also be perfectly accurate for a now-obsolete question. If the customer changes the product or requested remedy while a tool runs, evaluate whether the response is associated with the correct version of the task. This is more specific than asking whether the agent can remember a conversation.
A proposed factorial evaluation
Start with recorded, consented or purpose-created calls covering supported languages, realistic accents, interruptions, poor microphones, and long tool delays. Pair each call with a task state and expected external outcome. Use a controlled tool service so its responses and timing are reproducible.
| Controlled change | Comparison | Primary observation |
|---|---|---|
| Model version | Nova 2 Sonic versus 2.5 with the same application | Correct resolution, tool arguments, clarification rate |
| Connection renewal | Same script with and without renewal | Lost facts, repeated questions, delay after renewal |
| Tool timing | Result before, during, or after renewal | Missing, duplicated, or stale result handling |
| Acoustic processing | Echo cancellation enabled versus disabled | Self-interruption and intelligibility under speaker playback |
| Input representation | Original audio versus transcript replay | Errors hidden by text-only evaluation |
These are proposed ablations, not reported results. Do not combine every change in one upgrade trial: a new model, new audio processing, and new reconnection policy can improve the total experience while leaving the cause unknown.
Place renewals at ordinary pauses and at difficult moments. Include a correction immediately before renewal, an interruption during playback, and a tool timeout after submission. Observe attempted and completed tool operations independently of what the agent says. A spoken success claim should agree with the tool’s authoritative result.
Evaluate long calls separately from short calls. Report resolved-task rate, repeated-question count, entity retention, tool correctness, user interruption success, and tail response delay. Human listeners should judge naturalness without knowing the model, while task correctness is scored against the controlled environment. Neither score substitutes for the other.
The economic unit is a resolved call
The launch’s unchanged pricing claim does not imply unchanged call cost. A more capable model might resolve a call faster, ask more useful questions, or generate longer explanations. Renewal may carry additional context. Tool calls, telephony, recording, telemetry, and human handoff also contribute to total spend.
Use a proposed denominator of calls resolved correctly within an agreed observation window. Count callbacks and repeated contacts for the same problem. A lower initial handling time can be a false economy if the customer must call again after an incomplete resolution.
For a pilot, compare total operating cost and staff time per resolved issue, grouped by call type and language. Keep a separate measure for user effort: repeated facts, unexplained waits, and failed interruptions. A system can reduce labor expense while making the customer do more work.
Adoption is justified when the measured gains persist through realistic durations, acoustic conditions, and external tool delays. Nova’s model upgrade and Bidi’s conversation machinery make that test worth running. Their release labels cannot supply its result.
Dated source ledger and evidence limits
All sources were read on October 19, 2026. No service account or framework example was executed.
- Nova 2.5 Sonic launch — October 5, 2026; release, geography, relative pricing, and the 256K statement.
- Strands Bidi GA — October 5, 2026; connection lifecycle, audio processing, tools, and observability documentation.
- Nova 2.5 Sonic service card — scoped to the October 5 release; intended behavior and evaluation limitations.
- Nova models — living product page; the conflicting 1M statement, not proof of a dated context expansion.