Engineering
Can agents build and deliver software reliably?
The ability to plan, implement, review and ship software through issues, branches, worktrees, tests and recoverable orchestration.
Local MVP for a software factory
Cultivate, with evidence and limits.
Cultivate explains the move from one coding assistant to a Gas City software factory, then maps the capabilities needed for autonomous product organisations that operate under human authority.
- AssistantA model helps plan and write code.
- Agent teamIssues, branches and PRs let several agents work safely.
- Software factoryGas City keeps persistent work, formulas and sessions.
- Product organisationAgents test hypotheses before commissioning code.
- Institutional researchOrganisations and governance become research subjects.
Three parallel pillars
Product
Can agents discover what is worth building?
The ability to define outcomes, gather evidence, compare hypotheses, run experiments and decide when not to build.
Organisation
Can agents organise themselves to achieve outcomes?
The ability to hold roles, delegated authority, budgets, review duties and governance boundaries across teams and institutions.
Current state
What is working
The engineering workflow has issue claims, isolated worktrees, pull requests, automated checks and a running Mayor in Gas City. Slop Filter is the main verified example.
What is being tested next
The next product milestone is a bounded experiment comparing classification and exception approaches before commissioning more Slop Filter work.
Current frontier
The factory can deliver approved software. It has not yet proved that agents can decide which software is worth building. That is the shift from engineering automation to product organisation.
Browser verification is required before delivery claims
The process distinguishes synthetic checks, temporary Chromium release checks and installed signed-in Brave checks.
See this milestoneOutcome and baseline are written down
The first product experiment needs agreed outcome measures, representative-enough examples and a held-out evaluation set.
See this milestoneAuthority boundaries are explicit
Agents can recommend work, but new spending, public deployment, sensitive data collection and external actions require human approval.
See this milestoneCapability epics
E1
Coordinated coding agents
Agents can take approved engineering tasks, work in isolated branches and expose their progress through issues and pull requests.
Open epicE2
Persistent software factory
A configured Mayor and Gas City runtime can keep work, dependencies and sessions alive across interruptions.
Open epicE3
Verified continuous delivery
Software changes are not treated as delivered until the right automated and browser checks have run and their limits are stated.
Open epicE4
Economic and self-improving factory
The factory can reason about cost, capacity, quality and its own improvement opportunities.
Open epicP1
Outcome measurement and evaluation
The product layer defines useful outcomes and tests changes against evidence that can falsify weak ideas.
Open epicP2
Competing hypotheses and experiments
Several teams can propose different approaches to the same outcome, and an evaluator can compare them before build work starts.
Open epic