Claude Code dynamic workflows synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Claude Code dynamic workflows move multi-agent orchestration into a JavaScript script. The foreground Claude writes task-specific orchestration, or a command selects an existing script. A background runtime executes the loop, branching and fan-out; model-backed workers interpret tasks and use tools. Intermediate results remain in script variables, and the script-selected final answer returns to the foreground session. This is a division of decision ownership and context, not a claim that model reasoning becomes deterministic. See RT-BAP-script-control, RT-BAP-agent-prompts and RT-RTE-agent-orchestration.
The evidence is the single workflows documentation capture registered as SRC-1, frozen as workflows-2026-10-14 with an applicability cutoff of 2026-10-14. This is a doc-grounded, subsystem-only workflow analysis, not a whole-system assessment of Claude Code. The entire captured page was inspected by the source analysts. Actual workflow bodies, linked specifications, runtime implementation and execution evidence are outside the supplied evidence. Positive mechanisms below are documented contracts: implementation, activation and benefit remain uninspected. The boundary and Source register establish that scope. Both report verification and profile verification retain no blockers or additional declared limits. This synthesis uses their accepted records without additional source reads.
From request to admitted orchestration
A task can enter through an explicit request, bundled command, saved or plugin workflow, or enabled ultracode selection. In the ordinary manual CLI route, the user reviews phases, can inspect or edit the script, and approves or cancels launch. That review is not universal: auto mode normally records first-launch consent, skips the prompt with ultracode, and bypass mode launches immediately. Headless and SDK calls have no interactive prompt but still enter the enclosing permission evaluation. Restrictions on human-origin keyword triggers do not prohibit Workflow calls on other surfaces. Launch permission allows execution; it does not endorse worker results or approve every subsequent edit. See RT-RTE-workflow-launch.
The runtime's documented script isolation bars direct filesystem/shell access and module loading; workers perform tool actions under session permissions. Consequently, isolated orchestration does not imply isolated project effects. JavaScript determines which worker runs next and which returned values become subsequent inputs. Worker judgment remains model-mediated, and requested worker models can be substituted under organizational configuration. Neither exact model versions nor provider parameter changes are established. See RT-RTE-agent-orchestration and RT-CMP-claude-model.
Research, checking and delivery
The representative connected workflow is /deep-research <question>, selected because the page specifies a request-to-report progression rather than merely suggesting a task. It requires WebSearch and explicit invocation. The documented sequence searches across several angles, fetches sources, cross-checks and votes on claims, then presents a cited report in the session. The script owns orchestration; the precise worker assignments, handoff payloads and vote rules are unknown. The separate audit-script example cannot establish those research internals. See the workflow account, RT-CLM-research-report and EPI-RTE-research-acquisition.
Research checking and report disposition have different functions. Verifiers assess source support; the report process consumes those results to omit claims that fail cross-checking. When an API error or rate limit prevents checking, the claim is marked unverified rather than refuted. Thus report inclusion is not uniform verification. Surviving cross-checking is the intended inclusion criterion for the invoked question, but no actual evidence-consuming acceptance event, evaluator independence or calibrated threshold is supplied. Citations and votes do not establish an independent reference answer, and imported source warrant remains unknown. See EPI-RTE-research-check and EPI-RTE-report-disposition.
Other checks have narrower licenses. A supplied schema is checked for provable contradictions before worker launch; invalid structured outputs are retried and eventually fail. This admits structurally conforming values, not correct factual findings. The audit example's truthiness filter removes null slots without checking surviving findings or showing every discovered file was audited. Worker semantic transformations cannot be classified as faithful reshaping, entailed derivation or ampliative conjecture without concrete inputs and outputs. See EPI-RTE-schema-check, EPI-RTE-schema-disposition and EPI-RTE-result-filter.
Delivery means the finished workflow's selected report appears in the session. It does not mean independent human acceptance, complete question coverage or factual correctness. The operator can separately inspect prompts, results and partial tool-call displays through /workflows. A run cannot solicit stage sign-off internally; the documented alternative is separate workflows for separate stages. See the delivery and monitoring account. The claim of greater trustworthiness than a single pass remains a design assertion, not a measured comparison, in RT-CLM-research-report.
Failure, replay and reuse
Recovery changes what later execution receives. A stalled worker restarts with the original prompt but without its previous transcript; prior edits and spent tokens remain. Exhausted stalled calls inside fan-out continue with null, while a directly awaited stalled call ends the run with an error. Because the research script is unavailable, neither outcome can be assigned to every research stage, and an unverified-claim label does not promise report delivery after every failure. See RT-RTE-agent-recovery.
Relaunch requests recovery of retained results. The runtime reuses a completed prefix in original start order, then reruns from the first failed or changed-prompt slot, including later completed workers. Missing saved results produce an error; a fresh session starts a new run. Local results are session-associated, while cloud results can survive VM reclamation through conversation-history persistence. This is selective reuse, not rollback or exactly-once effects. Prompt compatibility also does not establish freshness after external files, sources, models or options change. See RT-OBJ-agent-results and RT-RTE-run-replay.
Saving and revising scripts preserves orchestration and embedded task guidance for later named invocations. Human saves and edits differ from runtime automatic persistence and model-supplied revisions. Loader, syntax and save-path checks do not establish a better method. In particular, project save protection checks ancestor and target symlinks, while personal protection checks only the target. Neither path establishes retained criticism or automatic learning from traces. See RT-OBJ-workflow-script and RT-RTE-script-admission.
Prefix caching is a different reuse mechanism. Matching model, effort, agent type, tools, schema and working directory permits later worker requests to reuse a tools/system prefix. The documented bounded hold and TTL concern processing reuse, not sharing sibling findings or transcripts. Persisted settings separately supply later consent, launch controls and authoring advice. See MEM-OBJ-prompt-cache, MEM-RTE-prefix-reuse and MEM-OBJ-workflow-settings.
Scenario-relative assessment and improvement boundary
For a task expressible as discovery followed by many independent worker calls, the distinguishing contract is script-owned fan-out with intermediate values kept outside foreground orchestration context. For research, it adds a declared check-to-report process that distinguishes failed checks from inability to check. For interrupted work, replay provides a documented way to reuse completed outputs, but repeated effects and stale inputs remain consequential alternatives. Resource ceilings limit agent counts and list sizes, not per-agent spend, wall-clock time or audit completeness. Size guidelines are advice, and disabling launch surfaces does not stop an existing run. See RT-RTE-agent-orchestration, RT-CLM-runtime-limits and RT-RTE-workflow-launch.
The strongest supported future-use contribution is retained orchestration and task-output reuse, not demonstrated improvement. Scripts can be revised and consumed later; results can replace repeated worker execution. What remains unresolved is whether evidence from operation shapes a durable change that improves future capacity. The requested checker/fixer loop concerns an external project and configured type-check success, not established improvement of the workflow system itself. See RT-RTE-script-admission, RT-RTE-run-replay and EPI-RTE-checker-loop.
The records do not establish theory-builder conditions 1–4, criticism of a consumed theory improving future capacity, reflection, autonomy or self-improvement at this boundary. These remain independent unestablished properties, not a combined negative verdict. Script editability establishes neither formulated theory criticism nor its iteration; automatic scheduling establishes neither every internal role's autonomy nor reliable model compliance. No trace-fed durable behavior-shaping derivation is established, but unresolved context and checkpoint alternatives also prevent a bounded absence conclusion. See RT-RTE-script-admission, the epistemic bounded conclusion and profile verification.
Inspectable orchestration and permission code would change the implementation assessment. Provenance-linked traces would establish which handoffs, checks and recovery branches actually operated. A controlled single-pass comparison could test the claimed research benefit; independently varied verification could support finer attribution. For improvement, evidence would need to connect an objective-bearing evaluation to an update of retained orchestration, its later live consumer and subsequent behavior depending on that update. Saving, loading or replay alone does not establish that chain.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Documentation without implementation or execution | SRC-1; RT-RTE-agent-orchestration, RT-RTE-workflow-launch, RT-CLM-runtime-limits | Entire workflows-2026-10-14 capture; linked specifications and host internals excluded | Actual enforcement, successful activation, deployment safety, reliability, performance and later-release applicability | Version-pinned runtime/host implementation, applicable configurations and provenance-linked executions |
| Research internals and comparative benefit unresolved | SRC-1; RT-CLM-research-report, EPI-RTE-research-check, EPI-RTE-report-disposition | Declared bundled procedure; actual script and run absent | Complete evidence delivery, independent verification, exact acceptance thresholds, factual acceptance and superiority over a single pass | Bundled script and prompts, claim/source/check traces, reference outcomes and controlled comparisons |
| Worker semantic provenance and coverage incomplete | SRC-1; RT-OBJ-agent-results, EPI-RTE-research-acquisition, EPI-RTE-result-filter | Generic worker contracts and example filtering, not actual task pairs | Preserved source warrant, warranted derivation, novel accepted claims or exhaustive audit/report coverage | Concrete inputs and outputs with source provenance, task inventory and accounted-for failed slots |
| Replay validity and persistence details incomplete | SRC-1; RT-RTE-run-replay, RT-OBJ-agent-results | Prompt/order/failure rules and local/cloud persistence contracts | Freshness under model/options/tools/file/source changes, atomic persistence, physical cleanup, null/error retention and exactly-once effects | Matching/invalidation and storage implementation plus change, interruption and effect traces |
| Context and checkpoint alternatives unresolved | SRC-1; RT-RTE-agent-orchestration, RT-RTE-agent-recovery, RT-OBJ-agent-results | Intermediate isolation, partial progress display and stall transcript discard; usage-wait context unspecified | Complete worker context, raw trace retention, summary/reload routes and complete memory inventory or trace-learning absence | Context-construction and checkpoint contracts/code, retained payloads and later-consumer traces |
| Prefix payload and control lifecycle opaque | SRC-1; MEM-OBJ-prompt-cache, MEM-RTE-prefix-reuse, MEM-OBJ-workflow-settings | Delivered prefix, matching/TTL/hold and documented settings consumption | Complete cache substrate/encoding/provenance/eviction classification, administrative backing and consent revocation guarantees | Provider cache specification/implementation and settings admission, persistence and revocation paths |
| Durable criticism and improvement chain unestablished | SRC-1; RT-RTE-script-admission, EPI-RTE-checker-loop, RT-CLM-research-report | Save/edit/reload and requested diagnostic feedback; actual scripts and authoring reference excluded | Retained revision/rejection reasons, consumed theory criticism, improved future capacity or exercised self-improvement | Evidence-responsive revisions with retained reasons, later consumers and comparisons tracing changed organization to subsequent behavior |
| Model internals and execution identity uninspected | SRC-1; RT-CMP-claude-model | Model selection/substitution documentation | Immutable executing model identity, parameter mutability or provider-level learning conclusions | Exact resolved model revisions and parameter/update contracts for the assessed operation |