Autogenesis synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From request to delegated action
Autogenesis is an enclosing runtime for user tasks, with a capability-improvement plane inside the same system. It schedules agents, calls remote models, dispatches actions and admits reusable extensions. This account covers the shipped implementation and callers at commit 5b070d5cd497e951a188a01409138e8b530eaf4a, with an analysis cutoff of 2026-10-15. SRC-1 is the sole frozen repository source. Code and shipped instructions support implementation and doctrine findings; no execution trace or causal experiment is registered, and the target was not executed. Remote providers, service internals and separately versioned dependencies remain outside the evidence boundary. See the boundary and evidence and Source register.
The ordinary path is client → Gateway task queue → MetaAgent → selected capabilities → returned result. Gateway binds session roots and grants; a conversation ID scopes model memory and budgets, while workspace resources remain session-shared. Runtime pumps deliver messages, Agent code reconstructs context and dispatches concurrent action batches, and a remote model chooses actions from projected schemas. The shipped direct driver bypasses Gateway but still invokes the included agent machinery. The configured model name is not an immutable weight identity: configuration and fallback can resolve later requests differently. RT-RTE-task-execution, RT-CMP-model-manager.
For substantial work, MetaAgent's prompt requests actor delegation, followed by verification and corrective rounds when needed. This is a source-supported workflow, not an observed episode or mandatory pipeline. Children receive a self-contained task, existing files and effective capability lists under a fresh context ID; they do not inherit the parent's whole conversation. Their successful message returns to the parent, not the complete trace or structured response data. Shared file effects remain possible. Blocked children can escalate through a bounded parent rendezvous, while action failures become corrective context. Selected skills take a different route: they return Markdown instructions to the same agent rather than launching another model or automatically executing scripts. RT-RTE-child-delegation, RT-RTE-skill-instructions, workflow account.
These distinctions matter for a delegated implementation or research task. The foreground window can stay small because child intermediates stay outside it, but each corrective or verification handoff must supply the evidence it needs. Neither context separation nor allowlists imply effect isolation. Bash checks its own registered permission mode and can launch a local shell when no peer sandbox is bound. Compiled workflows dispatch managers without the Agent's named-tool denylist, and direct playground chat bypasses the Agent task loop. Capability availability, session grants and deployed isolation are therefore separate boundaries. RT-RTE-shell-execution, RT-RTE-workflow-execution, RT-RTE-model-playground.
Carrying experience into the next decision
Autogenesis has a wired automatic trace-to-summary-to-later-model route. Hook-selected task and action records accumulate in recent memory; count overflow selects old records for a separate compacting model. Its short prose enters bounded working memory and later foreground context. This is a concrete trace-fed retained-context contribution, not merely log storage. It establishes neither improved future capacity nor faithful compression: exact-value preservation is requested, not checked. Empty summary output restores the chunk; exceptions after removal have only a logging branch. Count windows also do not impose a total prompt-token ceiling. MEM-OBJ-recent-history, MEM-OBJ-working-memory, MEM-RTE-tier-context.
The context builder requests history by context ID, then automatically supplies selected tails to the decision model without a model recall request. A separate TodoTool stores structured steps and Markdown, and context assembly rereads the whole matching Markdown even outside the memory-enabled branch. Tiered-memory JSON/HTML persistence has no corresponding restart-to-prompt loader in the inspected get/create path. Todo disk reread, transcript replay to clients and workflow recovery must not be generalized into universal Agent-history restoration. MEM-RTE-todo-context, MEM-OBJ-memory-display, MEM-RTE-client-replay.
Knowledge corpora provide another retained route: ingest appends caller documents and backend metadata; retrieval rereads them and uses BM25 or TF-IDF/cosine to return top-ranked records. These inspected rankers build transient statistical representations, not a retained neural embedding index. Workflow normalization preserves the returned data, but model delivery needs a downstream binding. Web acquisition can instead return model-generated summaries and merged reports directly. Neither source URLs, lexical relevance nor retention supplies a truth-acceptance decision; original source warrant remains unknown and report fidelity unverified. MEM-OBJ-corpus, EPI-RTE-corpus-ranking, EPI-RTE-web-reports.
Checking, changing capabilities and recovering
Checking has several owners and scopes. Model evaluator agents request test cases, source review and scored reports, which remain advisory. Workflow VERIFY returns one or several agent verdicts without aggregating them into acceptance. Exact-match benchmarks compute agreement with supplied references, not reference truth or a causal explanation of performance. Workflow evaluation records add stronger lineage controls: successful evidence must match a retained terminal run, version, program hash and recorded success; duplicate run IDs are rejected. Current-version summaries report healthy only with at least three distinct cases, success rate at least 0.8 and average caller-supplied quality at least 0.7. That signal informs inspection, not a mandatory keep/rollback branch. The immediate individual evaluation returned by evolution_tool need not equal the normalized stored record; later ledger reads are needed for stored run-derived metrics. EPI-RTE-evaluator-reports, EPI-RTE-workflow-checks, EPI-RTE-workflow-evidence, report-verification limits.
The strongest supported capability-change result is live and cross-restart admission machinery. A generator can write staged source, report its path and automatically trigger validation, promotion, dynamic loading, registration, archives and an active manifest. Later dispatch and startup loading can consume that retained implementation. Generated memory classes have a sibling hook, without demonstrated migration of existing sessions. Administrative configuration is another route and does not inherit every evolution gate. RT-RTE-capability-admission, RT-BAP-extension-execution, MEM-RTE-memory-admission, RT-RTE-runtime-configuration.
Two qualifications change what admission guarantees. First, the automatic tool-registration path consumes no per-content human approval despite prompt language requiring it. Second, the optional default-on smoke probe follows live registration and manifest writing; startup loading bypasses that probe branch. It tests general-agent success and absence of a constraint stop, not exact OK, required use of the changed capability, correctness or benefit. Failure attempts rollback/unload, but arbitrary import-time or external effects need not reverse. The evolution playbook also has a default skill-type projection mismatch, so its on-disk presence does not establish delivery to MetaAgent, although MetaAgent's own prompt contains evolution guidance. RT-CLM-user-approval, EPI-CLM-smoke-safety, RT-RTE-capability-admission, RT-RTE-skill-instructions.
Recovery is similarly specific. Workflow checkpoints check executable identity and reuse recorded completed invocations, but incomplete effects may repeat. Capability restore points contain version pointers rather than conversation state or complete system snapshots; individual restores can fail after others succeed. These mechanisms support bounded resumption and version selection, not exactly-once execution or transactional reversal of the whole system. RT-RTE-workflow-execution, MEM-RTE-version-recovery.
What completion and improvement mean here
For an ordinary open task, done_tool normally yields the final message and Gateway publishes completion on normal return. Review is requested model policy, not a hard terminal veto. MetaAgent can also terminate explicitly incomplete while setting the flag used by the common success calculation. Consequently delivery means return through the Gateway event channel, not independently verified goal satisfaction or user acceptance. For a predefined workflow, checks and numeric results can feed authored downstream control, but zero benchmark score alone does not reject a successful workflow response. RT-RTE-task-execution, workflow account, EPI-RTE-workflow-checks.
For improvement of reusable capabilities across later dispatches and restarts, the independent objective is better user-task completion. The implemented source/registry/manifest path supports standing improvement-directed self-change. It does not show objective-bearing evidence selecting a particular update or subsequent operation causally depending on that update. Exercised self-improvement, successful improvement and learning as improved future capacity remain unestablished; this does not negate the wired trace-fed summary update. RT-CLM-self-evolution, RT-RTE-capability-admission, MEM-RTE-tier-context.
For the localized evolution method, the runtime separately assesses theory-builder conditions 1–4: 1 afforded, because whole method units exist; 2 claimed, because content-dependent consumption is requested but unobserved; 3 claimed, because failure-directed criticism is requested without a supplied challenged claim, attribution and actual revision; 4 claimed, because repeated evaluation/revision is instructed without retained criticism shown shaping the next round. These statuses do not establish a theory builder. Reflection has inspection, feedback and intervention machinery, but no supplied bidirectional representation-mediated correction cycle. Computational authorship and automatic admission establish autonomy on those operations, not role-by-role autonomous theory building or reliability. No observed criticism of the method against its own operation establishes reflective theory building. These are independent findings, not stages in a single grade. RT-RTE-capability-admission.
A recorded episode connecting failure evidence, the chosen update, admission and later use would change the exercised-path assessment. Comparable outcome measurements could establish benefit at the comparison's actual scope; component-specific attribution needs independent variation of that component. Mandatory candidate exercise before admission, a real approval rendezvous, or a tier reload consumer would change particular architectural conclusions rather than merely add outcome evidence. The present records support none of those stronger conclusions. RT-RTE-capability-admission, MEM-RTE-tier-context, EPI-CLM-continuous-improvement.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Individual evaluation response can differ from stored normalization | SRC-1; MEM-RTE-evaluation-records, EPI-RTE-workflow-evidence; report-verification limits | Context-manager record copy versus evolution-tool serialization | Equality of submitted/returned individual metrics and stored run-derived metrics; ledger lineage checks and advisory summary remain supported | A later ledger read for the stored metrics, or a changed response path serializing the normalized record |
| Coarse diagnostic lineage classification omits supported local derivations | SRC-1; MEM-OBJ-conversation-events, MEM-OBJ-message-snapshots, MEM-OBJ-trajectories; profile-verification limits | Grouped diagnostic/caller-content lineage unit remains not-determinable despite automatic derivations in particular parts | Complete per-product lineage comparison for that group; existing system-level lineage positives remain valid | Separate classification of recorded observational derivations from unresolved caller content |
| No registered operation or causal comparison | SRC-1; RT-RTE-task-execution, RT-RTE-capability-admission, MEM-RTE-tier-context | Frozen implementation and doctrine only | Exercised task success, useful compaction, evidence-shaped capability change, causal benefit and improved future capacity | Inspectable task/update/use traces and objective-relevant comparisons with appropriate controls |
| Summary and imported-source warrant are unresolved | SRC-1; MEM-OBJ-working-memory, EPI-RTE-web-reports, MEM-OBJ-corpus | Inspected summary, merge and ingest/ranking routes | Faithful semantic preservation, justified external facts, or acceptance/integration of ampliative claims on these routes | Input/output pairs, warranted sources and evidence-consuming acceptance decisions for named uses |
| Recovery products have different consumers and horizons | SRC-1; MEM-RTE-tier-context, MEM-RTE-client-replay, MEM-RTE-version-recovery, RT-RTE-workflow-execution | Registry selection, disk displays/replay, version maps and workflow resume | Universal Agent-history restoration, complete rollback or exactly-once external effects | Traced missing reload/restore consumers and crash/recovery evidence with external-effect contracts |
| Included alternatives remain incompletely traced | SRC-1; memory integration issues, epistemic source-and-claim boundary, RT-RTE-workflow-execution | Custom memory/rankers, full frontend resend, feedback/export consumers, canvas/library, science/kernel and other evaluator leaves | Complete memory inventory, universal model-delivery or training claims, and system-wide absence of acceptance, derivation or theory building | Source-grounded producer/update/consumer and evaluator traces for the named alternatives |
| Admission and deployment checks are path-specific | SRC-1; RT-RTE-capability-admission, RT-RTE-runtime-configuration, RT-RTE-shell-execution | Generator hook, startup, administrative handlers and tool-owned shell effects | Universal human approval, candidate safety, transitive read-only confinement or complete side-effect reversal | Enforced gates covering material alternate paths, candidate-specific checks and deployment isolation evidence |
| Provider identity and remote behavior are outside inspected implementation | SRC-1; RT-CMP-model-manager, RT-RTE-model-playground, boundary evidence | Local clients/configuration, excluding provider internals and separately versioned dependencies | Immutable deployed weights, provider parameter-update claims, remote retention and actual service/isolation correctness | Pinned deployment/provider evidence and inspectable contracts or traces within an authorized expanded boundary |