Monorepo hosting as a self-healing control surface
All major infrastructure paths should stay close to the code, visible to agents, and increasingly capable of detecting, triaging, and fixing problems before humans are forced into manual loops.
Code-adjacent infrastructure
Runtime profiles, environment expectations, and service contracts live alongside the monorepo definitions that consume them.
Signal-native operations
Logs, traces, deploy events, and runtime health are promoted into machine-readable signals that agents can reason over continuously.
Agent-visible control plane
The more of the stack that is accessible through explicit APIs and policy-aware actions, the more proactive the platform becomes.
Hosting topology
Control plane
gRPC services, authz, orchestration state, event routing, and signal evaluation.
Object layer
S3/GCS compatible object storage with direct upload/download flows.
Coordination database
Postgres for refs, op-heads, presence, and upload sessions.
CLI transport surface
Native contract consumers inside local toolchains and agent workers.
Self-healing infrastructure
The end-state is not "better dashboards." It is infrastructure that can describe its own health in terms an agent can act on. When the runtime, logs, deploy system, and routing surface are all agent-accessible, the platform can move from passive observability to proactive operations.
Signals over raw noise
A signal is not just a log line. It is a structured conclusion drawn from frontend failures, API logs, health checks, release metadata, latency tails, and policy context.
Continuous triage
Agents should always be looking: clustering incidents, correlating them with recent deploys, identifying likely blast radius, and preparing the smallest safe fix.
Solutions prepared before execution
The platform should let an agent draft a rollback, patch a route, adjust traffic, or open a code change before a human even asks for help.
Operational model
Shared services, fewer blind spots
Keep a small number of shared services for identity, coordination, and caching so operational knowledge compounds instead of fragmenting per repo.
One control model across environments
The same control plane runs for local dev, CI, and production, with strict override profiles instead of separate stacks and dashboards.
Prepared actions, not ad hoc heroics
Rollback, restart, traffic shift, config patch, and hotfix flows should exist as first-class actions with audit trails and policy checks.
Levels of automation and control
Level 0
Observe
Collect signals, summarize probable causes, and notify humans. The system is agent-aware but does not change production.
Level 1
Recommend
Agents prepare a diagnosis and an exact remediation plan, but execution still requires explicit human approval.
Level 2
Guardrailed action
Low-risk playbooks can run automatically: restart unhealthy allocations, revert the latest release, disable a bad config, or shift traffic away from a failing path.
Level 3
Autonomous remediation
For well-understood failure modes, agents can patch, canary, verify, and self-revert if the signal does not clear, while humans retain policy and override authority.
Example signal path
A concrete example: the system sees the frontend failing on registration. Instead of waiting for a human to notice, VEX Source Control should be able to convert that into an operational loop:
01
The frontend begins emitting repeated `POST https://vex.sc/api/v1/auth/register` responses with `404`, or the registration path crosses a configured latency/error threshold.
02
That becomes a signal once it is sustained and correlated with recent frontend, API, or edge changes rather than treated as an isolated log line.
03
The agent inspects routing config, deploy history, logs, and ownership context to decide whether the issue is a stale client build, missing API route, or edge misconfiguration.
04
It prepares the safest fix first: restore the route, roll back the offending release, or ship a targeted patch behind a canary.
05
After execution, it keeps watching the same signal until the failure clears or escalates with a full incident narrative and proposed next action.
Hosting expectation
We are optimizing for predictable operations across environments: minimal moving parts, explicit deployment boundaries, reproducible infra assumptions, and a clear path from human-operated systems to supervised autonomy. The goal is not to remove operators. It is to give operators an agentic platform that is always watching, always triaging, and ready with prepared actions when the system drifts.