Engineering work and experiments

Problem

As agent workflows grow, useful engineering practice tends to become scattered across prompts, conversations, and harness-specific configuration.

Engineering ideas

• Portable procedure packages with explicit entry points and boundaries
• Composable workflows for planning, review, evaluation, context, and agent readiness
• Progressive disclosure and deterministic helpers instead of loading everything into context
• Durable workflow artefacts and explicit human responsibility boundaries

Problem

Skills need to be discoverable across different coding-agent environments without copying whole catalogues into context or losing track of the exact revision used.

Engineering ideas

• Intent-based discovery over approved skill repositories
• Immutable packages, digests, and lockfiles for reproducibility
• Harness-neutral materialisation rather than owning execution
• Revision-bound lifecycle observations and bounded feedback

Problem

Model routing often hard-codes provider names into the classifier or asks one expensive model to choose another, making policy difficult to inspect and evolve.

Engineering ideas

• Separate task classification from provider/model policy
• Small local classifier with explicit abstention and safe fallback
• Versioned routing registry for cost, latency, context, and modality constraints
• Reproducible training and privacy-safe routing telemetry

Problem

Long-running agent work needs durable state, bounded retries, independent verification, auditable effects, and recovery without requiring a large orchestration framework.

Engineering ideas

• SQLite as the durable state and audit ledger
• Coordinator-owned effects and explicit separation of powers
• Bounded specialist agents with structured attempts and escalation
• Recovery and reconciliation as first-class workflow states