After months of building autonomous AI agent pipelines, the patterns that matter most are not the ones I expected. The agents that deliver the most value are not the most clever — they’re the most reliable.
Structured Outputs are Non-Negotiable
Every agent in NightShiftCrew v2 must produce a validated artifact — code that runs, a dashboard with real computed metrics, a report grounded in filesystem data. Free-form text between agents is a recipe for cascading failures.
class AgentOutput(BaseModel):
status: Literal["success", "failure", "partial"]
artifact_path: str
exit_code: int
reasoning: strThis schema contract between agents is what makes the whole system reliable. The quality validator doesn’t use LLM judgment — it checks: does the code run to exit 0? That binary test is worth more than any subjective scoring system. The broader architectural principles behind this design are documented in the Multi-Agent Orchestration System case study and the earlier essay Multi-Agent Systems: Lessons from Production.
Local LLMs Changed the Economics
Running Ollama with Qwen 2.5 Coder 14B locally eliminated API costs and latency variability. When your agent loop needs 50+ iterations to converge on working code, paying per-token becomes untenable. Local inference made rapid iteration tractable. The same economic logic — that the constraint shapes the craft — is a theme I return to in On Finite Tokens and Infinite Tasks.
The goal was not “generate text about analytics” — it was “run analytics and show me the results.”
There is a deeper argument here about the distinction between generating text and actually doing the work, one that extends well beyond engineering. In Doing Academic Philosophy in the Age of AI, I explore the same structural limitation from the other direction: what happens when entire disciplines confront the fact that their most visible outputs were always the scaffolding around the real intellectual labor, not the labor itself.
Related
Projects: NightShiftCrew v2 — Autonomous AI Agent Pipeline · Multi-Agent Orchestration System · Semantic Document Processor
Writing: Multi-Agent Systems: Lessons from Production · On Architecture as a Design Discipline · On Finite Tokens and Infinite Tasks · Doing Academic Philosophy in the Age of AI
