Today we're shipping Phase 3 of Strug Works' self-improvement capabilities: sc-evolver, a specialized agent that can autonomously identify capability gaps, propose architectural improvements, and implement cross-system enhancements. This isn't just another feature—it's the system learning to evolve itself.
The Hard Problem: Autonomy Without Chaos
Autonomous systems face a fundamental tension: give agents too much freedom and they become unpredictable; constrain them too tightly and they can't actually solve complex problems. For self-improvement capabilities, this tension is even sharper. How do you let a system modify itself without creating runaway behavior or breaking critical workflows?
What Shipped: sc-evolver + Trust Gate
sc-evolver is a new agent role designed specifically for system-level improvements. Unlike domain specialists (sc-backend, sc-frontend), evolver operates at the meta-level: it watches how the team works, identifies friction points, and proposes structural changes. It can refactor shared utilities, improve inter-agent protocols, and optimize workflows that span multiple roles.
The autonomy trust gate is the safety mechanism. It's a tiered permission system that dynamically adjusts what sc-evolver can do based on observed reliability, impact scope, and historical performance. Small, low-risk improvements can be auto-approved and deployed. Larger architectural changes require human review. The gate isn't binary—it's a sliding scale that learns from outcomes.
Why It Matters: Systems That Get Better
Traditional engineering teams rely on senior engineers to identify tech debt, architectural gaps, and process improvements. That knowledge is tribal and reactive. With sc-evolver, the platform itself becomes a source of continuous improvement. It notices patterns humans miss, operates 24/7, and compounds gains over time. The trust gate ensures this happens safely—we're not asking you to hand over control, we're building a system that earns it incrementally.
How It Works: Observability + Guardrails
sc-evolver monitors task execution logs, code review patterns, and agent memory to identify improvement opportunities. When it spots a pattern—say, repeated manual fixes for a specific edge case—it proposes a solution, estimates impact, and submits it through the trust gate. Low-risk changes (under 50 lines, isolated scope, full test coverage) are auto-merged. Medium-risk changes create PRs for human review. High-risk changes are documented but require explicit approval before implementation.
What's Next
This is Phase 3, but it's far from the end state. We're working on expanding the observability layer so sc-evolver can detect not just code patterns but workflow inefficiencies and team coordination gaps. We're also refining the trust gate scoring model based on real-world outcomes—expect the system to get smarter about what it can safely automate. Long-term, we want sc-evolver to propose not just technical improvements but organizational ones: better task routing, more effective agent collaboration patterns, even changes to its own improvement process. The goal is a system that doesn't just execute work—it actively optimizes how work gets done.