An internal agent system in which natural-language seeds and end-to-end scenario harnesses drive agents that write, validate, and iterate on code under standing rules that humans neither write nor review code.
Natural-language seed through the validation-feedback loop to converged software.
Seed the work
A natural-language seed — a few sentences, a screenshot, or an existing codebase — starts the run[Source 2]
Implement and validate in the loop
Agents write code and run the end-to-end scenario harness; a sample of the output feeds back into the inputs so the system self-corrects[Source 1, Source 2]
Converge on the holdout scenarios
The loop runs until the holdout scenarios pass and stay passing, and the code ships without human review[Source 1, Source 2]
Where people stay involved
People review, approve, or guide the agent’s work here.
Reported: The standing rules exclude humans from writing or reviewing code; what happens when a run fails to converge is not documented, and the unknown attention-boundary judgment stays in research details.
Supervision level
Not classified
natural-language seed → agents implement and validate against scenario harness → converged software[Source 1, Source 2, Source 3]
Multi-provider routing by task as of 2026-06 (gpt-5.5 for everyday and DevOps tasks, QA orchestration, and security review; consensus of opus-4.8 and gpt-5.5 for sprint planning; opus-4.8 for frontend aesthetics and writing; gemini-3-flash-preview for image comprehension and agentic dialogues; gpt-image-2 and gpt-realtime-2 for UX ideation and voice)[Source 6]Observation date: 2026-06
Harness
End-to-end scenario harness with user-story-style scenarios kept outside the codebase as a holdout set and checked by an LLM; satisfaction measured as the fraction of observed scenario trajectories that likely satisfy the user; Digital Twin Universe clones of Okta, Jira, Slack, Google Docs, Google Drive, and Google Sheets absorb scenario load[Source 1, Source 3]
Sandbox
Not reported
Tool access
Filesystem that agents read and write to self-manage context; Digital Twin Universe as a stand-in for third-party services[Source 3]
Knowledge
CXDB, a self-hosted context store for AI agents with turn DAG, blob deduplication, dynamic types, and visual debugging[Source 4]
Context management
Pyramid Summaries with reversible summarization at multiple zoom levels; on-disk state as the memory substrate[Source 3]
Credentials
StrongDM ID, an identity system for humans, workloads, and AI agents with federated authentication and path-scoped sharing[Source 4]StrongDM ID covers identity for humans, workloads, and agents; secret handling inside runs is not further documented.
Interfaces
Not reportedNo human-facing surface is named; the factory is non-interactive and work starts from natural-language seeds.
Mechanisms
Digital Twin Universe
Behavioral clones of third-party services that replicate APIs, edge cases, and observable behaviors, so thousands of scenarios can run per hour without rate limits, abuse detection, or live-API cost[Source 1, Source 3]
Shift Work
Splitting work into interactive and fully specified classes; when specs, tests, and the existing app fully express intent, an agent runs end-to-end without back-and-forth[Source 3]
Attractor
Non-interactive coding agent structured as a graph of work phases with natural-language edges evaluated by the LLM; execution traverses the graph until convergence or termination, and the spec is open source[Source 5, Source 7]
Validation and failure handling
Scenarios
End-to-end user-story validations, often stored outside the codebase like a training holdout set, flexibly checked by an LLM to resist reward hacking and lazy test rewriting[Source 1]
Satisfaction
Probabilistic success measure replacing boolean test-green; the fraction of all observed scenario trajectories that likely satisfy the user[Source 1]
Reported observations
Not reported: The captures state principles, techniques, and model configuration; they report no measured outcomes of the factory's software. The token-spend figure is a stated operating principle kept under lessons.
Lessons
Reported opinion: StrongDM advocates evaluating generated code through observable behavior while treating its internal structure as opaque.[Source 3]
StrongDM describes a specification, validation harness, and feedback loop that stops when holdout scenarios pass and remain passing.[Source 2]
Reported opinion: StrongDM proposes spending at least $1,000 in tokens per human engineer per day as an operating benchmark for its software factory.[Source 1]
Reported: The standing rules exclude humans from writing or reviewing code; what happens when a run fails to converge is not documented, and the unknown attention-boundary judgment stays in research details.
implementation
Reported
validation
Reported: End-to-end holdout scenarios scored by satisfaction; the Digital Twin Universe carries the validation load and stays under implementation as a reusable mechanism.
observations
Not reported: The captures state principles, techniques, and model configuration; they report no measured outcomes of the factory's software. The token-spend figure is a stated operating principle kept under lessons.
No explicit confidence assessment is recorded for this claim.
SupportsSoftware Factories And The Agentic MomentDefinition: non-interactive development where specs + scenarios drive agents that write code, run harnesses, and converge without human review
The techniques page explicitly compares code with an opaque model snapshot; this is a stated evaluation philosophy, not proof that source review has no value.
The site documents zero human code-writing and review with automatic scenario validation, but no passage states that a person is summoned when scenarios fail to converge; what happens on non-convergence is undocumented.
