Maturity Path¶
I would build all of this in stages. Each stage adds one capability, and each stage's verification is what makes the next one safe to attempt.
| Stage | Name | Adds | Covered in |
|---|---|---|---|
| 1 | Tool-using engineer | Human → agent → tools. The agent can understand, edit, build and test | The Agent Runtime |
| 2 | Closed-loop engineer | Goal → change → verify → repair → verify | Verification |
| 3 | Planning | Goal → decomposition into work items → execute → verify the goal | Work as a DAG, Five Primitives |
| 4 | Persistent repository intelligence | Architecture, domain, requirement and historical memory | Memory & Knowledge |
| 5 | Parallel execution | Task DAG + change surfaces + write leases + worktrees | Ownership, Integration |
| 6 | Specialized reasoning | Architecture, security, testing and domain agents | Roles & Review |
| 7 | Portfolio-scale modernization | Many repositories, many objectives, central policy and knowledge, distributed execution | Reference Architecture |
Stage 7 is where this becomes transformative:
100 repositories
1,000 modernization objectives
hundreds of agents
central policy
central knowledge
distributed execution
Reading the path¶
Stages 1–2 are about one agent being trustworthy. Nothing later works without a closed verification loop; skipping to parallelism with an agent that can't verify its own work just produces unverified work faster.
Stage 3 is where the work item arrives. It is the smallest primitive and the one everything else attaches to — which is why it is the thing to build first.
Stage 5 is where the rest of this site becomes load-bearing. Before it, one agent works at a time and ownership is trivial. From it onwards, semantic conflicts, write leases and the merge queue decide whether parallelism helps or hurts — and your architecture decides how much parallelism is available at all.
Stage 6 comes late on purpose. Splitting one agent into specialists is only worth it once you can see a bottleneck the split removes — see Don't start with multi-agent. The agent hierarchy in the modernization overview lives here, not at stage 1.
Stages are cumulative, not sequential projects
Stage 4's knowledge model makes stage 2's verification smarter; stage 5's leases make stage 3's plans executable in parallel. Treat the path as an order of adoption, and keep investing in the earlier stages as you add later ones.