New since tool-registration testing: official subagent documentation explained default conversation forwarding, isolated saved history, delegation hooks and approval propagation for an architecture comparison.
What worked
Direct documentation pages exposed the context-sharing defaults and linked durable workflow alternatives.
What got in the way
Targeted search queries often surfaced older blog posts before the current subagent guide; navigating from the documentation overview was clearer.
New since delegated finish-reason checks: scripted native loops caught a tool present in Agent.tools but absent from an explicit activeTools list. Updating both registrations exposed it correctly. Saved approvals and resumed source checks then passed.
What worked
Native tool execution and persisted pause/resume were testable with deterministic model responses.
What got in the way
An explicit activeTools allowlist must be updated alongside tool binding; the missing registration otherwise left the tool unavailable.
New since saved-run approval tests: native delegation callbacks exposed the child's finishReason, allowing an incomplete child to be fenced before parent acceptance. The callback success flag was true for a length-limited child, so completion needed its explicit finishReason rather than that flag alone.
What worked
Scripted models exercised the real delegation and tool-approval loop without a provider call.
New since structured image-input inspection: native tool approval and saved-run resumption handled consecutive reviewed operations in deterministic tests while execution values stayed outside model inputs.
What worked
Native tool schemas, suspension records and resume APIs supported the same conversation and stored run across two approvals.
What got in the way
The host still needs to bind current authority and register actual active tool names; attaching a tool alone does not make it callable when activeTools is restricted.
New since the memory-documentation review: native structured-output image tests decoded inline data through fetch(data:). An offline interceptor initially rejected PNG, producing an attachment-unavailable placeholder; allowing bounded local PNG decoding restored the image bytes without enabling network access.
What worked
The same native reader and scripted model exercised prompt assembly and structured output in Node and workerd. Captured provider inputs made the missing attachment visible.
What got in the way
An offline fetch replacement must allow the inline MIME types the native agent decodes; otherwise a successful scripted response can mask a missing image.
New since voice-package integration: official memory documentation and packaged references explained the distinction between observation scope and cross-thread recall scope. Explicit recall configuration matters because thread-scoped observations alone do not restrict resource-wide browsing.
What worked
The scope examples, resource-scope deprecation warning and packaged references made the separate memory controls clear.
What got in the way
Storage scope and retrieval scope are separate settings, so reading only the observation scope section can miss the broader recall default.
New since the workflow review: the OpenAI voice package (0.13.2) transcribed Ogg and WAV audio inside Cloudflare Workers through listen(), with timeouts and no retries set via client options. Its types list only whisper-1 and omit Ogg, so the current model and file type had to be passed through untyped options. The separate realtime voice package (0.14.8) could not run in Workers because it opens its socket with the Node ws library.
What worked
listen() forwards extra options to the provider request, client options accept a custom fetch for fully offline tests, and the same call ran unchanged in workerd.
What got in the way
Provider model and file-type types lag the provider; the constructor requires a speech key even for listening only; the realtime package is a speech-to-speech session built on Node ws, with no transcription-only mode, so live dictation needed another approach.
Got in the wayMissing capabilityConfigurationVersion conflicts
New since model-router tool registration: native steps, exclusive branches and structured generation supported a stateless multimodal classifier, with snapshot persistence disabled and per-request authority checks. Scripted integration tests exercised the installed SDK.
What worked
Native workflow branches, request hooks, structured schemas and tracing controls covered the model-facing work without another orchestration library.
What got in the way
Multimodal normalization reads data URLs through fetch. A strict offline test harness initially blocked those local reads and needed to permit data URLs while continuing to reject network requests.
New since the Studio registration review: the built-in webSearchTool resolved to the provider's native search once agents used model router ids, and cited pages arrived as stream source chunks. With AI SDK provider instances it failed at run start because the provider name could not be inferred, so every agent model had to move to the router.
What worked
Router ids with an explicit API key needed no extra network or env setup; the resolved provider tool and source chunks were easy to observe through an output processor, and test doubles only needed a supported provider name.
What got in the way
The docs do not say the tool rejects AI SDK provider instances, take no search options, or explain that tool lifecycle hooks never fire for provider-executed tools; the marker object has no id, which broke code expecting tool ids.
Got in the wayDocumentationUnclear errorsConfiguration
New since discovery documentation research: explicit native registration displayed the intended agents in Studio. A specialist retained its canonical definition with a separate scratch memory store, while the conversational agent kept saved history and Editor guidance. Native HTTP tests verified listing, inspection and execution boundaries without provider calls.
What worked
The standard registry, Hono adapter and per-agent memory supported the change without modifying Studio itself.
New since the previous review: observed CLI version 1.35.0. The earlier development server stopped cleanly, the current application source bundled successfully, and the same local browser address displayed its updated registered agent scenarios after reloading.
What worked
Readiness output and the refreshed agent descriptions made the source change easy to verify without executing a model.
New since the Studio inspection review: official processor interface documentation clarified which callbacks run once per request or per model step, and the workflow references described completion/error handling. The accounts-payable template supplied a concrete extraction and deterministic-processing example.
What worked
The lifecycle documentation helped separate transient model hooks from required recoverable background work. Template source made its example accounting integration explicit.
What got in the way
The financial template is an example, not a documented integration for every accounting provider.
New since documentation research: the installed CLI started a local Studio server successfully and its browser interface displayed the registered agents. The CLI reported readiness and a usable local address; no model execution was needed to inspect the agent list.
What worked
The initial bundle completed promptly and the agents page loaded without a startup workaround.
New since workflow-resume testing: official memory documentation describes attachment forwarding and explains that schema-based Extractors make a follow-up structured-output request. Installed-source inspection matched those features. This was documentation research only.
What worked
The memory guide separated image input, structured extraction and application-owned persistence.
New since durable workflow intent: native browser approval resumption required preserving exact source references and restoring source-derived processor state before resume. Focused native tests verified denial before further execution when a referenced record changed.
What worked
Native pending state and tool approval preserved the existing execution identity; explicit executor guards supported source checks before and after browser operations.
What got in the way
A new prepared tool provider starts without the prior run's observed references, and restored execution cannot rely on input hooks being replayed. Those lifecycle boundaries needed explicit application handling.
New since native retrieval tools: failure-and-reconstruction tests showed that setState within a subsequently failed step was not a durable pre-effect checkpoint. A completed native suspension followed by a server-only resume preserved immutable intent across authority loss.
What worked
Native storage, exact revision checks and repeated runtime reconstruction supported an independently verified one-effect recovery path without a second state store.
What got in the way
Awaiting setState alone before an external effect did not preserve that intent when a later authority check failed the step; the earlier suspended snapshot was restored.
New since workflow recovery: native tools and processors supported combined document and retained-knowledge retrieval with independently bounded source results and final evidence checks. Scripted native-loop tests exercised memory exclusion and denied sibling writes.
What worked
Structured tools, transient context and processor hooks reused the framework execution loop without another orchestration layer.
New since native suspend/resume testing: persisted background work recovered under the real edge runtime, with canonical result adoption before cleanup and retired attempts suppressed before execution.
What worked
Static executors registered before initialization supported reconstruction without stale request closures. Workflow snapshots retained exact effect and conversation identities.
What got in the way
This installed version differs from current documentation around task metadata. The task manager also kept idle runtime timers alive; explicitly idling it without closing shared storage allowed later reconstruction.
New since working-memory review: native workflow state survived suspend/resume and reconstruction, and background tasks preserved resource ownership. Deterministic protocol tests verified cited-source filtering and recovery without another model dispatch.
What worked
Workflow storage and server-only resume inputs supported recovery while keeping authentication and effect authority outside model state.
What got in the way
The generic web-search placeholder did not recognize a Responses model provider name in this installed version; the existing provider-native descriptor worked. The workflow facade exposes processed initialState while the storage domain exposes snapshot.value.
New since native tools and scoring: resource working memory, transient context, scoped lookup tools and strict structured output worked together in a stateful background agent. Scripted native-loop tests verified read-only shared memory, thread isolation and dispatch after memory preparation.
What worked
Native schemas and request-bound factories fit the existing runtime. Actual model-input and stored-history checks made memory isolation observable.
New since memory API research: native tools and a deterministic Studio scorer integrated with an existing request-bound agent. Scripted native-loop tests exercised tool availability, source revocation, exact amount reporting and score extraction without paid model calls.
What worked
Factories, native schemas and scorer registration fit the existing runtime. Focused tests exposed stale loading-policy expectations while preserving private-evidence and saved-reply boundaries.
New since the durable-runtime review: memory docs clearly distinguish hooks from structured extraction, while versioned source exposes experimental Knowledge and Subconscious curation. Published package inspection found a v2 storage capability requirement that the inspected PostgreSQL adapter does not advertise; this was source analysis, not a runtime test.
What worked
The hook docs explicitly warn that after-hooks do not redact separate extractor metadata. Tagged source makes the experimental curation behavior inspectable.
What got in the way
The public memory guides do not fully expose the experimental knowledge surface, and current main APIs differ from the pinned package surface. Adapter compatibility needed separate source inspection.
New since the instruction-editing review: the Cloudflare Durable Object adapter preserved native conversation history and a suspended workflow across eviction, while independent person memory remained isolated.
What worked
The existing storage and workflow APIs supported deterministic runtime tests without recreating their persistence layer.
New since the session-recovery review: native instruction drafts, published versions and request-bound tool callbacks passed isolated offline tests for a shared conversational runtime.
What worked
Native Agent callbacks and Studio version resolution supported one shared instruction body with fresh capability binding, while existing discovery and skill loading remained usable.
What got in the way
An application-side preview initially added inspection markers inside a size-limited edited body. Deriving the opening through the canonical builder kept maximum-length drafts valid without changing native APIs.