Autogenesis boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

Autogenesis is in scope as an enclosing runtime: it owns agent execution, issues model calls, dispatches capabilities, and serves user tasks through shipped drivers and interfaces. Its intended use is task completion through a MetaAgent and other actors, with generator, optimizer and evaluator agents for changing the capability ecosystem. The improvement plane is part of this target, not a separate external application. This is a whole-system functional boundary over the shipped Autogenesis implementation at the reviewed commit, including its direct example drivers, Gateway, browser/terminal clients, retained state and extension lifecycle. It is not a boundary around only the runtime Python module or only the MetaAgent API. SRC-1: README.md, pyproject.toml, autogenesis/cli.py, examples/run_meta_agent.py, autogenesis/agent/types.py, autogenesis/extension/server.py.

The classification rests on executable ownership, not the project description alone. Runtime spawning creates and registers an agent pump (SRC-1: autogenesis/runtime/server.py):

ref = AgentRef(name=ref_name, agent_name=agent_name, status=AgentStatus.RUNNING) ref._pump_task = asyncio.create_task(_pump(agent, ref), name=f"pump-{ref_name}")

self._refs[ref_name] = ref --- autogenesis/runtime/server.py

The agent decision path streams a model with messages and callable capabilities (SRC-1: autogenesis/agent/types.py):

model_manager.stream( name=self.model_name, input={"messages": messages, "tools": tools}, ctx=ctx, ) --- autogenesis/agent/types.py

Entry points and state-loop coverage

The installed console entry is autogenesis.cli:main. Its command and terminal modes create an in-process Gateway; serve mode exposes the Gateway through stdio or WebSocket transport. Gateway startup initializes capability managers, extensions, task workers and restored sessions. The direct MetaAgent example instead bootstraps managers itself, binds session roots, installs a TaskManager handler and invokes agent_manager. That shipped caller is included even though it bypasses the Gateway. SRC-1: pyproject.toml, autogenesis/cli.py, autogenesis/gateway/service.py, examples/run_meta_agent.py.

The inspected state loop includes creation and consumption of per-session memory, persisted conversations and sessions, workflow checkpoint recovery, persisted knowledge corpora, and reusable extension admission. These are distinct stores; writing a memory report does not by itself establish that a restarted agent reloads it. The inspected tiered-memory get reads an in-process session registry, whereas Gateway session recovery reads manifests and knowledge retrieval rereads corpus JSONL. Later analysis must distinguish those paths rather than generalize persistence into universal restart recovery. SRC-1: autogenesis/memory/default/tiered.py, autogenesis/agent/types.py, autogenesis/gateway/service.py, autogenesis/knowledge/server.py, autogenesis/workflow/runtime.py.

The extension consumer is inside the target. Startup reloads the active extension manifest, and the Gateway promotion handler registers promoted components through add_component. The inspected implementation archives versions, writes the manifest and can revert after a failed smoke probe. Shipped evolution instructions and prompt text are also included as declared policy; they are not assumed to be identical to enforcement. In particular, the self-evolution playbook describes register-is-live behavior, while the MetaAgent prompt also describes session staging and user-approved promotion. This boundary retains both surfaces so later jobs can trace their respective callers rather than silently choose one description. SRC-1: autogenesis/extension/server.py, autogenesis/gateway/service.py, autogenesis/skill/creator/self_evolving_skill/SKILL.md, autogenesis/prompt/default/meta_agent.html.

The reusable-state reload responsibility is explicit in the extension implementation:

await self._load_component(comp.module, abspath, comp.name, version=comp.version, config=None) self._ensure_archived(comp.module, comp.name, abspath, comp.version) loaded.append(comp) --- autogenesis/extension/server.py

Evaluation is included on three different surfaces: model-driven evaluator agents, benchmark adapters and smoke-probe admission checks, plus repository regression tests as executable specifications. None of these source files is an observed successful run. The default smoke probe calls a general-purpose agent on a synthetic completion task; its presence does not establish that an evolved capability is actually exercised or improves task quality. SRC-1: autogenesis/agent/evaluator/agent_evaluate_agent.py, autogenesis/benchmark/server.py, autogenesis/extension/smoke_gate.py, tests/test_runtime.py.

Repository coverage

The tracked root tree was enumerated at the pinned commit. “Included” below specifies analytical responsibility, not a claim that every implementation file was read during boundary declaration. Initial inspection traced the material entry, state production, admission, reload and consumption paths above. Uninspected leaf implementations remain available within SRC-1; they are not access gaps and cannot support detailed conclusions until inspected.

Frozen evidence and external dependencies

The sole frozen evidence source is SRC-1 at full commit 5b070d5cd497e951a188a01409138e8b530eaf4a. The supplied checkout's HEAD matched that pin and its short status was empty. The source object above is unchanged from acquisition. The analysis cutoff is 2026-10-15, the supplied run date; it does not assert repository completeness after the reviewed commit. No branch tip, live website, second repository, historical analysis or newly executed target output is evidence for this run.

Overall evidence tier is code-grounded. Inspected Python, shell and client call sites supply implementation evidence; target READMEs, prompt text and skill instructions supply doctrine/design. No observed-run or causal-experiment source is registered. The target was not executed, installed or tested. Consequently this boundary supports source-level responsibility and mechanism analysis, not claims of real-world success rates, reliable self-improvement, measured retrieval quality, operational security or causal benefit.

External dependencies include model-provider APIs/SDKs, MCP connector services and plugin-backed services, Docker/container and browser infrastructure, Vault or environment-supplied credentials, package dependencies and optional benchmark data. Their shipped adapters, configuration and launch contracts are included; remote internals, credentials, live service responses and separately versioned dependency implementations are excluded. These exclusions prevent conclusions about provider behavior, connector service correctness, actual deployed isolation, credential configuration and dataset availability. Nested Gitlinks do not authorize inspecting another repository under SRC-1. SRC-1: pyproject.toml, README.md, autogenesis/gateway/service.py, docker/base/Dockerfile, scripts/serve-ui.sh.

Source register