Who Are We Now?

The piece argues that software knowledge comes in two forms: the visceral, participatory understanding gained by making and changing code (Verum Factum), and the objective, experimental knowledge gained by testing and observing behavior (Bacon’s vexations). It frames a running system as a symmathesy - a learning system composed of people, code, tests, logs, CI and observability - where confidence to ship comes from being deeply involved. As teams scale, individual ownership weakens but collective verum factum can be rebuilt by increasing learning flows: pairing, expressive domain code, thorough tests, and observability so the system teaches developers and developers teach the system.

Introducing agents and LLMs changes the balance: an agent cannot genuinely “make” and be changed by its work, so it lacks verum factum and instead thrives on rigorous experimentation. That requires a tenfold investment in objective verification - unit/property tests, CI/CD, per-change verification, observability, evals for LLM-driven components - and automated checks and scaffolding that reduce vigilance toil. Agents participate only through the context and tools developers provide, so developers must build and tune that world, set goals, and take on architect/product roles. Practical consequences include more test and instrumentation code than production code, continuous automation of checks, and redefining legacy as code not set up for agents.