Speed without governance creates risk. Governance without speed creates irrelevance. CAS (Coding-Autopilot-System) is built on the position that this is a false trade-off: when governance is executable — schemas, gates, and evidence instead of meeting minutes — it is the speed.
This page is the organization-facing version of the flagship vision document. It exists so anyone landing on github.com/Coding-Autopilot-System can read the thesis and reach every repository behind it without cloning anything first.
Every enterprise is discovering the same thing: AI agents can produce code far faster than humans can review it. The standard responses both fail —
- Ship it (speed without governance): untracked artifacts, untruthful commit messages, unreviewed merges, silent drift. We have measured versions of each of these failure classes in our own workspace — including a commit that claimed to add test suites while its diff contained none.
- Board it (governance without speed): review queues, manual sign-off, documentation that decays the day it's written. The agents idle; the advantage evaporates.
CAS is a working system — 13 repositories in this organization plus a workstation control plane — that resolves the tension by making every governance control machine-executable and evidence-producing.
flowchart LR
subgraph Control["Control plane - gsd-orchestrator (.NET 10)"]
G[Goal admission<br/>typed, bounded, budgeted] --> S[Dependency-aware<br/>scheduling]
end
subgraph Execution["Execution plane - autogen / Microsoft Agent Framework (Python 3.12)"]
S --> W1[Worker fan-out<br/>isolated mutation scope]
W1 --> T[Typed failure states<br/>structured JSON telemetry]
end
subgraph Governance["Governance plane"]
P[Promptimprover<br/>prompt & learning governance] -.gates.-> W1
C[cas-contracts<br/>versioned JSON schemas] -.validates.-> T
E[cas-evals<br/>deterministic verification] --> V{Evidence<br/>gate}
end
T --> E
V -->|proven| M[Merge via PR<br/>branch-protected, two-party]
V -->|failed| R[Bounded repair<br/>or deterministic halt]
R --> S
Three planes, strictly separated, each owned by a repository in this organization:
| Plane | Owner | What it enforces |
|---|---|---|
| Control | gsd-orchestrator |
Goals are typed, measurable, bounded. Nothing runs without a budget and a stop rule. |
| Execution | autogen (MAF workers) |
Every worker owns its mutation scope exclusively. Failures are typed states, never stack-trace soup. |
| Governance | Promptimprover, cas-contracts, cas-evals |
Prompts, schemas, and verification are versioned artifacts. Completion requires deterministic evidence — an agent saying "done" is not evidence. |
flowchart TD
A[Plan<br/>phase contract with threat model] --> B[Execute<br/>agent fan-out, one owner per repo]
B --> C[Test<br/>ratcheted branch-coverage gates]
C --> D[Verify<br/>goal-backward evidence check]
D --> E[Review<br/>PR, branch protection, two-party approval]
E --> F[Audit<br/>milestone verification vs falsifiable requirements]
F -->|drift found| A
C -.->|coverage regression| B
D -.->|gaps| B
The loop is not aspirational — each gate is a running control with a falsifiable check:
- Truthful history: a commit-integrity check flags any
test:-typed commit whose diff contains no test files (built after catching a real one). - Ratcheted coverage: CI gates fail on branch-coverage regression and only ratchet upward with real tests — never a hardcoded aspirational 100%.
- Typed failures: a versioned
FailureStateschema incas-contracts; orchestrator and workers map every boundary failure to it, proven by fault injection (process kills, corrupted checkpoints, simulated API outages). - Supply-chain pinning: every third-party GitHub Action pinned to a commit SHA with least-privilege token scopes, verified by an org-wide lint that runs in CI.
- Drift detection: a daily workspace-health sweep runs 11 checks (unpushed work, orphaned artifacts, stale PRs, encoding hazards, broken worktrees) and exits non-zero on any finding.
- Deploy lock as policy: infrastructure is maintained "bicep-ready" — linted, parameterized, pinned — while an operator hard lock prevents any cloud deployment until it is deliberately lifted. Governance includes what you don't ship.
flowchart TD
O[Orchestrator<br/>frontier model, full context] --> PL[Planner agents<br/>mid-tier]
O --> EX[Executor agents<br/>mid-tier, one repo each]
O --> VR[Verifier / critic agents<br/>mid-tier]
O --> DOC[Doc & housekeeping agents<br/>small models]
PL -->|plans as contracts| EX
EX -->|evidence| VR
style O fill:#1f6feb,color:#fff
The orchestrator delegates through typed sub-agents (planner, executor, verifier, critic, doc-writer), each with a scoped toolset and an explicitly chosen model tier — frontier reasoning only where it pays, small models for mechanical work. Specialist agents deliberately cannot spawn their own sub-agents; delegation depth is a governance decision, not an accident. The result: parallel throughput of a team, token cost of a task queue, and an audit trail for every decision.
- 25-PR release train merged across 13 repos with branch protection never left disabled.
- 2 orphaned test suites recovered and gated; coverage baselines measured and ratcheted (branch coverage, not vanity line metrics).
- 37 + 16 new tests landed via fault injection and coverage phases; 134-test suites green.
- 4 stranded commits, 11 unmerged agent branches, and 14 broken worktrees detected and recovered — zero work lost.
- Every finding became a permanent automated check the same week it was found.
Every plane above is backed by a real, buildable repository in this organization. Start wherever matches your interest:
| Repository | Role | Plane / category |
|---|---|---|
| gsd-orchestrator | Autonomous issue-to-PR engine; goal admission and scheduling | Control |
| autogen | Multi-agent worker fan-out on Microsoft Agent Framework | Execution |
| Promptimprover | MCP-first prompt governance and traceability | Governance |
| cas-contracts | Versioned lifecycle schemas (prompts, events, artifacts, FailureState) |
Governance |
| cas-evals | Deterministic evaluation and evidence-gate verification | Governance |
| cas-reference-product | End-to-end reference workload with a verified case study | Proof |
| cas-platform | Azure hosting/observability foundation — bicep-ready, deploy-locked | Platform |
| cloud-security-service-model | Enterprise cloud security operating model | Platform |
| cas-workstation | Windows-first AI-native developer workstation bootstrap | Workstation |
| autopilot-core | Control plane for org-level CI repair automation | CI Automation |
| ci-autopilot | Worker/runtime for queued repair execution on self-hosted runners | CI Automation |
| autopilot-demo | Bounded demo target for the full failure-to-fix loop | CI Automation |
| .github (this repo) | Org governance, community health files, and this profile | Org |
For a shorter front-door version of this page, see the organization profile README.
This page is adapted from the canonical docs/VISION.md, maintained at the root of the CAS
workstation workspace — the source of truth for this thesis, tone, and diagram style. See each
repository's own docs/wiki/ for implementation-level detail. Diagrams are Mermaid-first;
codex:generate-image placeholders mark where generated visuals will replace them.