All work
AI engineering · Internal R&D
AI Agent & Automation Experiments
Controlled experiments orchestrating coding and operations agents with explicit human approval gates.
Prototype / ExperimentPublic summary — client anonymised
Context & challenge
Understanding where AI agents genuinely help engineering and operations work — and where they need hard limits.
Objectives
- Evaluate agent orchestration on real engineering tasks
- Define approval gates for risky actions
- Compare local and hosted models
Solution approach
- Agents with scoped tools and repositories
- Board-style approval rules for deploys, merges and credential changes
- Local LLMs for private, low-risk tasks
Architecture
- Task
- Orchestrator
- Scoped agents
- Tool calls
- Human approval
- Repository / system
Simplified and anonymised. No private infrastructure details are published.
Important technical decisions
- Destructive and production actions always require explicit human approval
Operational considerations
- Credential handling
- Agent drift from scope
- Auditability of actions
Results
Results pending verification. We only publish outcomes that have been confirmed.
Lessons learned
- Agent usefulness is bounded by the clarity of its scope and its approval rules.