Everything around the model that decides what it can see, what it can do, and when it is done. It is the file the agent loads first, the phase boundaries that decide which context comes next, the skills and references it can reach for, the guides that prevent wrong moves, the sensors that catch them, and the artifacts that record what was learned. Same model + better harness = different system.
Synthesis of Mitchell Hashimoto (HashiCorp), Birgitta Böckeler & Martin Fowler (Thoughtworks), and the Anthropic Engineering blog — popularised under the name harness engineering by BettaTech, ¿Qué es esto del Harness Engineering?, April 2026.
The harness assumed the agent could head a CSV. Production is streaming topics, joined-at-query-time, schema-drifting, secrets-gated.
The Environment box stops being a folder.
10 kHz CAN bus, MB-per-row sensor frames, waveforms, point clouds. No LLM has seen the bytes — skills do the reduction and hand back distilled views.
Skills get heavier; tool outputs need budgets.
MATLAB, Simulink, CAD, FEA solvers — license-gated, GUI-bound, no MCP wrapper. The tools chip hides a $40k/seat license behind FlexLM.
tools is not a Python function — it's a fleet of integrations.
Read-only on prod, write on staging, secrets via vault. Identity-as-agent ≠ identity-as-user; every tool call carries a who / what / why envelope.
The tools chip becomes a permission boundary.
Skills are load-bearing code now. PR review on SKILL.md, an owner per skill, versioning so an agent mid-run doesn't read today's skill against yesterday's plan.
The ratchet needs a code-review process attached.
TREE.json survives one run. Production needs the org-scale version — why we tried X, why we abandoned Y, what our "we tried that already" memory looks like.
Registries grow from per-run files to a shared substrate.
Regulated industries need every decision traceable to a human-readable why — in a log that survives a model upgrade re-running the task differently.
Artifacts stop being internal memory and start being evidence.
Same model + better harness = different system. Production needs more harness.
Each item above is a new component, a new sensor, or a new guide on the diagram you just saw.