JAZ synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From request to native result
JAZ is an enclosing runtime for model-directed Python computation. A host supplies named instructions, data and callable capabilities; JAZ queries a model, executes its response as Python, returns observations and repeats until native return or raise. Its advertised iterative computation is supported as implemented capacity, not measured task success. The evidence is the generated Python distribution at commit 0803d4971be785e95b80054b02259664d70fa3da, applicable through 2026-10-15. The whole-system boundary includes the shipped console, recursive calls, hooks, execution controls and retained-execution tooling. SRC-1 supplies implementation and design text, but no execution trace or causal experiment. Provider internals, model weights, an external trainer and unshipped evaluation artifacts are not frozen evidence. See the boundary, Source register, RT-CLM-repl-work and RT-RTE-agent-loop.
Library invoke/ainvoke and the console reach the same Agent-owned loop. Defaults are LiteLLM, PythonREPL and CodeOnlyProtocol, but a model must be selected and components are replaceable. Each invocation creates fresh messages and a live Python namespace. The model chooses source; the runtime schedules query, parse and execution. CodeOnlyProtocol treats the response as code rather than extracting a fenced block. Ordinary execution errors become feedback for another turn; terminal values and exceptions pass through completion hooks before reaching the caller. There is no inherent iteration cap or ordinary expected-answer oracle. See RT-CMP-model-backend, RT-OBJ-message-buffer, RT-OBJ-repl-state and RT-RTE-agent-loop.
Recursive calls delegate selected arguments and inherited scope to a fresh Agent, then return native results or exceptions. They do not automatically copy the parent transcript. Fresh contexts reduce transcript exposure, but shared host references prevent an object- or process-isolation interpretation. Input/output abbreviation and hook edits separately control the visible window. The default limits are per-input/per-output character ceilings, not a total token reservation. Full Continue output remains available through REPL history even when its displayed observation is shortened; terminal history fields are not equivalent checkpoints. Persistent message edits persist within the invocation, not automatically across calls. See RT-RTE-recursive-work, MEM-RTE-context-selection and EPI-RTE-observation-cut.
Controls govern actions, not a universal truth criterion
Hooks can modify inputs, messages, source and results, supply responses or results instead of running ordinary producers, or abort. Consequently, receiving a result does not prove a model query or local execution occurred. Optional code predicates can veto committed source before execution; return predicates can reject an already-produced value and feed criticism back into the current run. ReturnType checks type conformity. None supplies general task correctness without a substantive caller criterion. BudgetForcing merely refuses early finishing for a configured number of attempts. See RT-RTE-hook-control, EPI-RTE-code-validation, EPI-RTE-return-validation and EPI-RTE-budget-forcing.
These controls have conditional force. Ordinary hook bugs are logged and swallowed rather than universally failing closed. Optional budget checks are not reservation-based hard monetary ceilings; the recursion cap has a documented wrapper/thread ancestry gap. Agent-source import/file restrictions do not recursively constrain host callbacks: the shipped SWE library can run shell commands and modify real files if granted, though it is not the default grant. Resource guards do not establish OS confinement. This architecture supports controlled in-process computation under trusted host components, not an unconditional sandbox guarantee. See RT-RTE-hook-control, RT-RTE-recursive-work and RT-RTE-host-tools.
Execution supplies evidence of what a computation did under particular bindings, not automatically of premise truth, explanation or transfer. Likewise, successful type checks, operational continuation and human consent have different licenses. No concrete supplied answer establishes either warranted derivation or a tested and accepted ampliative claim. The accepted account therefore distinguishes executable authority from epistemic warrant rather than grading JAZ as a whole. See RT-BAP-response-execution, EPI-RTE-return-validation and the epistemic conclusion.
The console connects delegation to delivery
The representative multi-step workflow is a % request to the shipped settings helper. The console delegates proposal generation with a redacted configuration snapshot, API reference, source-read access and text transcripts—not the live console namespace. A restricted, recursion-leaf helper returns a typed answer/code tuple. Type failures can cause another turn, but this is not semantic snippet review. Initial turn exhaustion can prompt fresh attempts with caps of 10, 20 and 40 after the initial 3; these restart rather than resume the failed helper. Session-budget exhaustion does not receive more calls from those offers. See RT-RTE-settings-change and the workflow account.
Answers print without action confirmation. Code is displayed and requires explicit interactive approval; non-TTY execution is refused. Approved code runs unsandboxed in the host namespace, not inside the restricted helper. The recorded applied=True flag precedes execution and means admission, not successful completion. Only a returning executor reaches the success marker; failures can leave partial effects without rollback. This separates proposal, consent and delivery. Ordinary console Python and other host capabilities are outside this helper gate. See RT-BAP-helper-proposal and workflow delivery.
What survives and shapes later work
The console automatically retains requests, live input/result references and exceptions across prompt turns. Later main invocations receive both a capped transcript display and the live history object, which generated code can inspect. Helpers receive text transcripts only. Per-entry display caps do not bound turn count, guarantee relevance or serialize arbitrary payloads. Rejected helper snippets remain context without acquiring execution permission. See MEM-OBJ-conversation-history, MEM-OBJ-helper-history and MEM-RTE-session-history.
Longer-lived reuse takes several forms. ATIF recording retains conversation and metadata, including failed runs, but serializes arbitrary values as capped representations rather than restartable Python objects. Replay loads caller-selected traces and automatically supplies saved responses into the normal code-execution path. It can repeat effects. Its shared-stack selector corresponds to the querying invocation only under strictly sequential nesting; concurrent invocations can receive another frame's code or divergence state. Divergence detection defaults off and, when enabled, does not compare the seed prompt or undo prior effects. Replay is therefore executable reuse, not a certified safe cache. See MEM-OBJ-atif-trace, MEM-RTE-trace-recording and RT-RTE-trajectory-replay.
The strongest supported trace-fed contribution is automatic production of durable workflow programs from execution-associated code and feedback, with an afforded standalone executor. This supports a bounded trace-learning affordance: experience can become a later behavior-shaping program. It does not establish successful execution, semantic equivalence or improved capacity. Failed/aborted workflows can be emitted, code-edit hooks can create capture mismatches, and automatic installation as the next Agent policy is not established. Diagnostic Markdown/log conversion has a different, interpretive consumer. See RT-RTE-workflow-emission, MEM-RTE-trace-display and the profile verification.
Training support is another partial contribution. Token rollouts preserve backend records, including hidden reasoning omitted from visible text, and export masked samples. A token-less completed query withdraws that invocation's rollout; nonmonotone histories refuse flat export but permit other projections. These are representation controls, not reward or truth checks. SGLang also exposes weight-update transport, but no accepted evidence joins sample production, optimizer admission, durable parameters and later generation into an evaluated learning loop. See MEM-OBJ-rollout, MEM-RTE-rollout-export and MEM-RTE-weight-transport.
Console defaults can separately persist and control later startup. The model-switch probe tests one minimal provider request before this particular switch; direct configuration and startup flags bypass it. It establishes neither future availability nor competence. Configuration revision is not itself learning. See MEM-OBJ-user-settings, EPI-RTE-model-ping and EPI-CLM-model-validation.
Assessment and what would change it
For iterative computation with host-chosen tools and validators, the inspected loop supplies explicit execution, recovery and return seams. For later reliance on saved actions, safe correspondence instead depends on sequential replay, unchanged relevant conditions and external effect management. For optimization, the supported contribution is reusable programs, token samples and update interfaces—not demonstrated better future action. The reported code-framing comparison remains an attributed, task-dependent robustness argument, not independently verified causal improvement or a general task-score advantage. See RT-RTE-agent-loop, RT-RTE-trajectory-replay, EPI-CLM-agent-optimization and EPI-CLM-code-framing.
The records support only a partial account of the independent theory and improvement properties. For theory-builder condition 1, configuration snapshots and source are inspectable content, but the records do not establish a formulated theory meeting that condition. For condition 2, delivery and instructions to consult that content do not demonstrate reliable content-dependent theory use. Conditions 3 and 4—formulated content-directed criticism and its retained influence on another round—are not demonstrated. Generic validator feedback is not enough to establish them. This is not evidence that opaque model processing contains no theories. See RT-RTE-settings-change, EPI-RTE-return-validation and the epistemic conclusion.
Criticism of a consumed theory has not been shown to improve future capacity. The helper has a wired configuration-representation and approved-change path, but the records do not establish reliable reflective use. Computational scheduling and proposal generation are implemented, while the selected action workflow retains human consent; autonomous theory building or optimization is not demonstrated. Exercised self-improvement is independently unestablished: no objective-linked evidence-to-update chain is traced through changes to JAZ's own organization and subsequent operation causally depending on them. Workflow artifacts may remain external products, and weight requests do not prove server adoption. These unresolved properties do not erase the supported trace-to-program contribution. See RT-RTE-settings-change, RT-RTE-workflow-emission, MEM-RTE-weight-transport and the epistemic conclusion.
Concrete evidence that would alter these assessments includes paired capture/execution traces showing generated-workflow equivalence under code edits and object binding; invoke-ID-safe replay selection tested with overlapping calls and changed seed inputs; and an inspectable trainer/update consumer linking an independently stated objective, update decision, parameter identity and subsequent behavior. A theory-builder finding would additionally need the localized theory, its actual consumer, criticism naming the challenged content and blame, and the retained consequence in a next round. These are missing evidential links, not a requirement that JAZ adopt one particular knowledge representation.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Unresolved implementation-status scope conflict | SRC-1; MEM-RTE-weight-transport, EPI-RTE-weight-update; retained by report verification | SGLang exposed methods and outbound HTTP wrapper | Memory assigns afforded, epistemic assigns wired to the unified mechanism at different scopes. No single unqualified implementation-status comparison is warranted; neither establishes a wired learning loop. |
Agreed separation of exposed capability, request wiring, caller activation and server application in the declaring records. |
| Static source rather than supplied execution or causal evidence | SRC-1; RT-RTE-agent-loop, RT-RTE-settings-change, EPI-CLM-code-framing | Frozen generated distribution; attributed experimental comments, without raw runs | Successful delivery, exercised recovery, reliable compliance and causal benefit cannot be inferred from implementation or reported comparisons. | Inspectable runs with task criteria and results; intervention comparisons with appropriate attribution limits. |
| External and opaque retained state remains unresolved | SRC-1; MEM-OBJ-conversation-history, RT-CMP-model-backend, MEM-RTE-rollout-export, MEM-RTE-weight-transport | Known containers/displays, sample exports and server-request interfaces; opaque live payloads and external trainer/provider internals | No complete memory inventory, exact learned-policy provenance, trainer-to-parameter chain or improved future capacity. | Payload and parameter identities, external persistence/admission contracts, training decisions and traced later consumers. |
| Replay correspondence and safety are conditional | SRC-1; RT-RTE-trajectory-replay | Shared-stack/cursor selection and optional signature comparison | No per-invoke correspondence under concurrency, seed-equivalence guarantee or rollback of repeated effects; positional selection also does not establish identifier-matched push. | Invoke-specific selection and seed/condition validation, with overlapping-call traces and explicit effect recovery. |
| Retained artifacts do not certify equivalent or successful successors | SRC-1; RT-OBJ-repl-state, MEM-OBJ-atif-trace, RT-RTE-workflow-emission, MEM-RTE-rollout-export | Continue-specific history, capped/shallow trace serialization, early workflow code capture and trusted token stamps | No full namespace restart, lossless terminal reconstruction, workflow equivalence, independent token authentication or quality-selected adoption. | Complete binding/state provenance, committed-code capture and paired executions; authenticated backend token/wire evidence and separate outcome criteria. |
| Selected audits do not cover every adapter or remote lifecycle | SRC-1; MEM-RTE-trace-display, RT-CMP-model-backend, RT-RTE-hook-control | Selected local converters, backend integrations and hook seams; not every malformed-input branch or telemetry consumer | No universal provider parity, complete remote-memory lifecycle or system-wide absence of additional evaluators/learning routes. | Scoped inspection of the remaining material adapters, transformations and actual external consumers. |