Autogenesis synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From request to model-directed work
Autogenesis is an enclosing multi-agent runtime with an internal capability-generation and optimization plane. It claims to complete user tasks through tools, delegated agents, skills and connectors, and to repair missing or weak capabilities. This account covers its shipped Gateway, direct Python drivers, workflow/canvas execution, retained state and extension admission at commit 5b070d5cd497e951a188a01409138e8b530eaf4a, with an analysis cutoff of 2026-10-19. SRC-1 supplies source-code implementation and prompt/design evidence. No target execution, task trace or causal experiment was supplied or performed. External services, dependency internals and excluded gitlinks are not frozen. The source register fixes those limits; neither source inspection nor the independent verifications establishes successful operation.
A Gateway submission returns task and conversation identities before execution. A worker invokes MetaAgent through a runtime that owns the agent reference, mailbox, outstanding actions and reply. Each round rebuilds context and capability schemas, then streams a configured model. The model chooses actions; code routes them, applies hooks and limits, executes capability managers and constructs the terminal response. The ordinary configured endpoint is google/gemini-3.1-pro-preview, not an immutable weight/build identity. Provider training and actual endpoint resolution remain uninspected. See RT-RTE-gateway-task and RT-CMP-model-manager.
Delegation gives a child a fresh context with selected task text, existing files and explicit restrictions or target defaults. The child's returned message crosses back, not its complete transcript. Shared workspace, services and registrations remain shared effects. Parent allowlists are not intersected with child grants on this path. A skill call instead returns method text and resource information to the calling model; it does not start a separate worker. These distinctions determine both evidence transfer and context size: separate work keeps internal child history out, but the parent must explicitly package requirements and evidence for subsequent workers. See RT-RTE-child-delegation, RT-RTE-skill-delivery and WF-RTE-ui-verification.
Dispatch rejects names outside the round's assembled routing table. That invariant does not confine arbitrary shell or generated Python effects. The agent's read-only deny set covers five named mutating tools, not Bash; Bash uses its own permission registration, and the default imported configuration permits full access. Without a bound peer sandbox it launches a local subprocess. No deployed isolation or exploit is demonstrated. Workflow steps call managers directly, direct model chat omits agent task/review gates, and ProceduralAgent runs a procedure rather than the model-round loop. Controls must therefore be stated per route, not as universal runtime guarantees. See RT-RTE-direct-agent, RT-RTE-workflow-execution and RT-RTE-model-chat.
Retaining experience for the next decision
The strongest supported experience-fed contribution is a complete automatic path from lifecycle records to retained summaries and later model input. Hooks project task/action/results into resident recent history. On overflow, code removes the oldest chunk, supplies it and the previous summary to a model, and stores any nonempty summary. Empty output restores the chunk. Later agent requests select recent records and working summaries by session identity and count windows. The summary also reaches the next summarization call. This is wired trace learning in the write-route sense: experience produces retained behavior-shaping material. It is not evidence of learning as improved capacity for future action. See MEM-RTE-history-compaction, MEM-RTE-tiered-context and the independent profile verification.
The selection is chronological rather than task-relevance ranking. Count limits do not bound tokens, and final-result text has a separate, unbounded-by-fetch-count inclusion. Nonemptiness is not a fidelity check. Successful compaction replaces raw detail without a per-summary assertion-to-event map; bounded summary storage eventually discards older chunks. Exact semantic preservation remains indeterminate without an input/output pair. The retained text supplies context, not verified facts or enforced conduct. See MEM-OBJ-working-memory, EPI-RTE-history-compact and EPI-RTE-summary-retain.
Persistence has distinct consumers. Tiered memory writes JSON/HTML displays, but its inspected built-in backends do not reload session tiers from disk after eviction or new-instance construction. Todo plans do reload JSON and supply same-ID Markdown automatically during context construction, independently of the memory switch. Conversation events reopen in the shipped client, not in a tiered-memory loader. Knowledge corpora reload on explicit queries and return lexically ranked document bodies in result data; a workflow must map those data to any later model. These are separate routes, not interchangeable evidence of cross-restart model recall. See MEM-ABS-tiered-reload, MEM-RTE-todo-context, MEM-RTE-conversation-replay and MEM-RTE-corpus-retrieval.
Checking a product and deciding to finish
The representative shipped request is a React shopping site whose acceptance conditions require working browser interactions after deployment. It is an intended workflow over implemented mechanisms, not a completed episode. MetaAgent is directed to delegate implementation, obtain the deployment's actual URL, pass that exact URL to a browser agent, and optionally send requirements, artifacts and browser evidence to a reviewer. Reported defects should lead to parent-directed repair, redeployment and renewed checking. A redeploy may change the URL and invalidate earlier evidence. The reviewer is instructed to recommend continuing when only curl or bundle evidence exists, but its verdict is advisory text. See WF-RTE-ui-verification.
This workflow distinguishes product repair from runtime self-change: fixing a cart or image does not itself improve Autogenesis's capability organization. It also distinguishes requested acceptance from runtime return. Deployment success means a running-status record, not demonstrated checkout behavior. Normal agent completion uses done_tool, but limits and no-progress stops can end work earlier. MetaAgent has a repeated-batch path that marks done while reporting incompleteness. Gateway's task.completed event therefore certifies a return, not successful storefront delivery. Browser/reviewer checking is not a mandatory generic completion gate. See RT-RTE-gateway-task.
There are nevertheless concrete narrow checks. Exact-match benchmarking computes equality against supplied references; it does not validate reference truth or identify the answer-producing mechanism. Workflow evaluation admits records only after version/run/hash/terminal-status consistency checks, requires retained real runs for successes, and rejects duplicate run evidence. Its health summary aggregates current-version evidence, but quality scores remain caller-supplied and the inspected consumer does not automatically roll back from that summary. These are distinct licenses: reference comparison, evidence-ledger admission and advisory assessment, not a universal factual acceptance process. See EPI-RTE-exact-match and EPI-RTE-workflow-evaluation. Generated factual novelty and accepted domain knowledge remain undetermined without proposition-bearing outputs linked to inputs and checks; this is not an absence claim about unassessed domain evaluators.
Changing capabilities and carrying changes forward
Autogenesis has implemented improvement-directed self-change capacity, beyond producing external work products. A successful generator or optimizer can trigger software validation/promotion, register changed capability source, retain archives and a manifest, and expose supported capabilities to later dispatch. The generic smoke check runs after live loading, is optional—including through an explicit false forwarded by Gateway—and need not invoke the changed component or compare its output to the requested OK text. It consumes response success and constraint status. Rejection can revert or unload registration, not undo arbitrary import-time or external effects. Prompt requests for explicit user approval are not an enforced approval rendezvous. See RT-RTE-extension-admission, RT-BAP-capability-load and RT-CLM-evolution-guidance.
Memory source admission has a consequential exception: class registration stores instance=None, so lifecycle hooks and prompt reads skip that configuration until a separate build or supplied-instance path provides a live instance. Archiving source and executing a constructor do not establish later memory delivery or session migration. Separately, the creator's renderer guidance conflicts with consumption: overriding _render changes persisted views, not inherited get selection. The memory evaluator also uses mismatched interface assumptions, and its named probes were not located in bounded searches. These findings qualify that specific improvement pathway without negating working built-in memory or other live capability changes. See MEM-RTE-memory-admission and MEM-CLM-memory-guidance.
Workflow checkpoints carry another bounded form of reuse. Resume checks runtime-major and workflow name/version, and rejects unequal nonempty program hashes. A supported hashless migration path instead stamps the current hash and can cache completed work without verifying its original executable identity. Neither branch establishes output truth or exactly-once external effects. Canvas JSON reuse and capability-version checkpoints are different retained artifacts, not generic agent mailbox or memory restoration. See RT-RTE-workflow-execution and EPI-OBJ-checkpoint.
The theory-builder findings remain independent. For the formulated evolution method, condition 1—localized content—is wired in prompt/playbook text. Condition 2—meaning-dependent consumption—is requested but unobserved. Condition 3—content-directed criticism—is prescribed, with no supplied challenged claim, result, blame assignment and revision episode. Condition 4—retained criticism shaping a next round—is likewise prescribed, not observed. The ordinary skill projection additionally filters out the orchestrator-tagged evolution playbook, although evolution guidance is directly present in the prompt. These limits do not erase the localized method or generator capabilities. See RT-RTE-extension-admission and RT-RTE-skill-delivery.
Criticism of a consumed theory improving future capacity is unestablished. Reflection is afforded by source/configuration representations that can guide changes to the represented capabilities; an exercised bidirectional episode is unobserved. The computational task/delegation/admission routes do not establish autonomous theory building: computation performing every criticism and retention operation has not been shown. Reflective theory building is also uninspected as an exercised property. For self-improvement, the boundary is Autogenesis's capability organization, the horizon is later calls/sessions, and the independent objective is successful user-task execution. Wiring supports improvement-directed change, but no supplied episode closes objective-bearing evidence shaping an update through later operation causally depending on it. Thus exercised self-improvement and successful improvement are not established, separately from wired trace learning and self-change capacity. See RT-RTE-extension-admission and MEM-CLM-memory-guidance.
For the shopping-site scenario, the supported assessment is capacity to orchestrate implementation and evidence handoffs, with model-owned acceptance and no demonstrated delivery. A linked browser/deployment/review trace would establish exercised verification; independent acceptance results would support bounded delivery claims. For capability improvement, an objective-bearing failure, selected revision, live registration, later invocation and controlled outcome comparison would change the present attribution. Building a usable memory instance, correcting renderer/reader guidance and aligning evaluator probes would repair specific downstream gaps, but would still not prove benefit. No target source was read for this synthesis; its account uses the supplied records. Both independent verifications declare no blockers or additional limits (report verification, profile verification).
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Source-only evidence; no linked execution or intervention | SRC-1; RT-RTE-gateway-task, RT-RTE-extension-admission, WF-RTE-ui-verification | Frozen implementation and design, not executed tasks | Observed compliance, delivered product correctness, exercised theory criticism/self-improvement and causal benefit | Linked task/update traces, independent acceptance results and comparisons at the claimed attribution level |
| External models, services and deployment conditions are not frozen | SRC-1; RT-CMP-model-manager, RT-RTE-direct-agent, MEM-OBJ-plugin-history | Local adapters/configuration, excluding service/dependency/gitlink internals | Immutable model identity, provider training verdicts, deployed isolation, service reliability and complete external history classification | Pinned provider/deployment identities, actual grants/mounts and inspectable service behavior |
| Memory content and provenance are not independently warranted | SRC-1; MEM-RTE-history-compaction, MEM-OBJ-working-memory, EPI-RTE-corpus-ingest | Projected lifecycle history, nonempty-summary admission and caller corpus import | Faithful summaries, preserved reasons/source warrant, accepted factual knowledge or improved capacity | Input/output pairs, provenance links, fidelity checks and downstream reliance/outcome evidence |
| Persistence routes have different consumers and horizons | SRC-1; MEM-ABS-tiered-reload, MEM-RTE-conversation-replay, MEM-RTE-todo-context | Built-in tiers, plan files and client transcript handlers; frontend not fully traced | Universal restart recall, complete transcript recovery or cascading deletion of all retained state | A tier loader or separately established backend, plus restart and deletion traces for each claimed consumer |
| Customized memory admission and guidance do not close the intended recall path | SRC-1; MEM-RTE-memory-admission, MEM-CLM-memory-guidance, MEM-OBJ-memory-display | Class-based extension registration, inherited get and renderer/evaluator contracts | Registration alone yields live memory; renderer changes improve prompt retention; evaluator scores validate retention | Explicit build/instance and migration wiring, corrected reader guidance, callable aligned probes and recall evidence |
| Operational checks do not certify substantive quality | SRC-1; RT-RTE-extension-admission, EPI-RTE-exact-match, EPI-RTE-workflow-evaluation | Optional post-load smoke, supplied-reference equality and evidence-ledger guards | Validated-before-live admission, mandatory human approval, verified quality or component-specific improvement | Pre-effect enforcement if claimed, target-specific probes, warranted references/quality measures and consequential selection traces |
| Hashless restore retains unverified original executable identity | SRC-1; RT-RTE-workflow-execution, EPI-OBJ-checkpoint | Supported hashless migration and nonempty-hash validation branches | Universal executable-identity assurance, fresh cached outputs or exactly-once external effects | Trusted original executable provenance and effect/freshness validation covering the restored work |
| Relevant alternative consumers and checks remain incompletely inspected | SRC-1; MEM-OBJ-execution-records, MEM-OBJ-plugin-history, MEM-RTE-corpus-retrieval, EPI-RTE-memory-evaluate, profile verification | Partial domain evaluators, process/workspace recall, provider histories, feedback and trajectory-training successors | Complete memory classifications, all-route negatives about checking/learning, trained successors or absence of accepted knowledge everywhere | Records tracing those specific transformations, admissions and later consumers; deployed evidence where implementation is external |