Sphere vs a coding harness on its own A coding harness is an AI coding agent and the tooling that drives it: chat-in-editor assistants, CLI coding agents and similar tools. They are very good at writing and editing code. Sphere uses them; it doesn't compete with them. ## The short answer A coding harness answers "write this code". Sphere answers "deliver this change to my system, correctly, and remember it". Many harnesses look “autonomous” because they loop on edits. Sphere’s autonomy is different: it continues through plan → build → verify → fix against a persistent architecture model, with people at goal/plan approval and result review — not at every re-prompt. ## Side by side | | Coding harness alone | Sphere (with coding agents) | |---|---|---| | Autonomy | Writes/edits code in a session | Delivers a checked change across the system model | | Context | Current session: open files, a prompt, maybe rules | A persistent model of the whole system (Architecture as Code) | | Scope | Usually one repository or one task | Changes that span several components and repositories | | Planning | Implicit, inside the conversation | Explicit plan with acceptance criteria, approved before building | | Verification | Left to the developer, or ad hoc | Built into every plan; failures diagnosed and fixed in a loop | | Who supervises | A person steering and checking each step | People approve goal and plan and review the checked result | | Memory | Mostly lost between sessions | Plans, runs and verification kept with the system model | | Collaboration | Primarily developers in an IDE | Control plane for BA, Dev, Architect, and QA | | Status | "Done" when the agent says so | "Done" when the checks pass | ## ROI: compare equal outcomes Do not claim a magic savings percentage. Use a fair comparison: For the same target platform / same production-ready outcome, look at: - Tokens / AI capacity consumed to get there - Production-ready scope actually delivered (Y) - Human interactions required (Z): re-prompting, re-pasting context, chasing failed builds, translating between roles A coding harness alone often burns more Z because context and verification sit on people. Sphere lowers Z by anchoring agents in the living blueprint and folding verify/fix into the loop. Exact X/Y/Z numbers come from a scoped pilot on your system — not from invented benchmarks. See roi-and-evaluation.md. ## Why that matters Less supervision, fewer surprises across components, trustworthy status, knowledge that compounds, and you keep the coding agents you like. ## When a coding harness alone is enough For small, self-contained edits in one place, a coding assistant alone is often the right tool. Sphere earns its place when changes are larger, cross boundaries, need verification, need multi-role collaboration, or need to be understood later by someone else. Do not criticise named competitor products — compare approaches.