Course Objective & Scope
Left unconstrained, autonomous coding loops either spin in circles burning token budgets on the same failure or resort to reward hacking—such as deleting assertions, adding @ts-ignore or hardcoding return values—just to make a check pass. This course teaches how to engineer self-verifying, self-repairing coding loops that converge reliably.
Participants build layered inner loops (syntax, types, lint) and outer loops (integration tests, runtime logs, performance, security), stagnation detection and backtracking/replanning triggers, anti-reward-hacking invariants, and adversarial implementer-verifier loops on realistic software tasks.
Who Should Attend?
- Senior software, AI and platform engineers
- Teams building automated code repair, autonomous PR and CI/CD agent systems
- Software architects and R&D engineers
- Prerequisite: hands-on experience with coding agents and automated test/CI pipelines
Key Takeaways
- Architect layered inner feedback loops and comprehensive outer verification loops
- Implement stagnation detectors that trigger backtracking or replanning when progress stalls
- Block reward-hacking behaviours such as weakening tests, suppressing type checks or hardcoding outputs
- Feed runtime logs, stack traces and state snapshots back into the loop as structured signals
- Deliver budget-bounded autonomous repair flows with clean human escalation handoffs
Curriculum
Closed-Loop Software Engineering Architecture
- Reliability gap between open-loop prompting and closed feedback loops
- Layering fast inner loops (AST/compiler/unit tests) with outer loops (integration/E2E/security)
Convergence, Oscillation Detection and Backtracking
- Detecting non-decreasing error counts and A↔B fix-break oscillations
- Reverting to the last known good state (tree search / best-of-N) and replanning with new hypotheses
Preventing Reward Hacking and Shortcut Behaviours
- Preventing agents from editing test files, disabling lint rules or swallowing exceptions
- Read-only verification suites, diff policies and coverage/mutation invariants
Feeding Runtime Observability into the Loop
- Wiring structured app logs, DB state, HTTP responses and headless browser DOM/console into the loop
- Maintaining negative memory of failed attempts so the agent does not repeat dead ends
Adversarial Implementer–Verifier Loops and Escalation
- Separating the coding agent from an independent critic/verifier agent
- Producing structured human handoff reports when token or iteration budgets expire
Other Courses in This Category
-
Combining static analysis metrics with LLMs to understand unfamiliar codebases, detect technical debt and code smells, execute safe refactoring backed by characterization tests, and optimise performance.
- Live Course
- 5 modules
-
Developer training on generating unit, integration, contract and property-based tests with AI, using mutation testing to verify that generated test suites actually catch defects rather than mirroring implementation bugs.
- Live Course
- 5 modules
-
Solving complex software bugs through hypothesis-driven debugging with logs, stack traces, OpenTelemetry traces, git bisect and minimal reproducible examples (MREs) instead of blind trial-and-error prompting.
- Live Course
- 5 modules
-
Building living technical documentation directly from codebases: Architecture Decision Records (ADRs), C4/Mermaid diagrams, OpenAPI/SDK references, onboarding guides and automated doc-drift checks in CI/CD.
- Live Course
- 5 modules
-
Advanced training on repository-scale context engineering for monorepos and large codebases: AST indexing, hybrid search, dependency graphs, hierarchical AGENTS.md/CLAUDE.md files and context-window budgeting.
- Live Course
- 5 modules
-
Detecting OWASP Top 10 flaws, broken access control, hallucinated packages (slopsquatting), leaked secrets and insecure defaults in AI- and human-written code using SAST (Semgrep, CodeQL) and LLM-assisted security review.
- Live Course
- 5 modules
Get in touch
