[2/2] The New SDLC: harness, factory model and agentic engineering economy (Category AI4SDLC)
I continue to analyze the May whitepaper from Google, where first We discussed the shift from syntax to intent, and then the spectrum of vibe coding → agentic engineering and context engineering. Now let's talk about what surrounds the model and turns it into a working agent.
Favorite formula: Agent = Model + Harness. When you start working with agents, it’s tempting to think of the system as a model: a new model comes out – the agent gets smarter, the old one is dumber. The authors say this is the wrong view, leading to the wrong investment. The model is just one entrance. Everything else—prompts, tools, context politics, hooks, sandboxes, sub-agents, observability—is harness, the strapping around the model that gets it done. According to him, the model is approximately 10%, harness - approximately 90Percentage of what you feel when working with Claude Code, Cursor, Codex or Gemini CLI
The evidence is not rhetorical. At the benchmark Terminal Bench 2.0 One team picked up a coding agent from outside the top30 top-down5Changed only harness, without changing the model. A separate study by LangChain raised the score 13.7 Tweaking only system prompt, tools and middleware around a fixed model. The conclusion that is worth hanging on the wall is that most agent failures, to be honest, are configuration failures, not models.
From this grows the factory model. **The main product of a developer is no longer code, but a system that produces code.**Specifications and context, executing agents, tests and quality gates, feedback loops that return errors to the agent, and guardrails. The plant manager does not sharpen every detail with his hands - he designs the conveyor and quality control. The modern developer gives agents success criteria, not step-by-step instructions, and lets them iterate.
The role of the developer is divided. 1е Conductor – real time, in IDE, keystroke control; good for research, prototype, learning a new API 2Orchestrator - asynchronously, at the level of goals, delegation to several agents, reviewing the result, not each line; good for features, migrations, test generation. Most switch between modes during the day.
Separately, “problem” 80%. The agent quickly generates around 80% of the feature code but the remaining 20% (edge cases, error handling, integration points, subtle correctness) It requires a deep context that models often lack. The nature of errors has changed: from syntactic to conceptual – wrong assumptions about business logic, missing edge cases. They are harder to spot precisely because the code “looks right” and passes basic tests.
The authors are interested in economics. For a leader, velocity is the total cost of ownership. Vibe coding looks cheap (capex)But hides high OpEx: token burn, constant request cycles, maintenance tax on spaghetti, security patching. According to the authors, at the intersection point of vibe coding costs in 3–10More for a feature. Agentic engineering flips model: high CapEx (Specks, tests, context structuring) And low margin OpEx. Context engineering and intelligent model routing (Big models -- complicated, cheap -- deterministic.) They become direct financial instruments.
The practical part of “where to start” is divided into three levels. 1Developer: start AGENTS.md, set a set of skills, make the first agent from repetitive workflow, write tests and evals before code generation, review every line that goes into the product. 2Leader: Make Context Engineering a Competent Practice (AGENTS.md, prompts, evals, skills - as code, under review and versioning n) And set the bar on eval, not demo. 3Organizations: invest in production substrate to scale, adopt open standards (MCP, A2A)Plan human/agent hybrid teams.
The document is titled “Intent as the New Interface” and has three principles: 1️⃣ Structure scales, vibes don't 2AI strengthens your engineering culture – multiplies both strengths and weaknesses 3The role of man evolves, not disappears Final sentence.
Generation is solved. Verification, judgment, and direction are the new craft
For me, this is exactly the logic that I understood in Agent-first IDP article: Agent = Model + Harness, evals as a quality contour, production substrate to scale - the same ideas, but at the platform level. The main conclusion of both texts is the same: discipline. (specs, tests, evals, harness) Not the opposite of speed, but its condition.
#AI #AI4SDLC #VibeCoding #Engineering #Architecture #Management