AI EngineerVideoSource Linked

fighting slop with slop — Vaibhav Gupta, Boundary

Boundary replaces blanket code review with stable architecture rules, agent-generated tests, transcript inspection, execution traces, and human triage—a concrete harness pattern, not a license to skip verification.

AI Engineer · Jul 31, 2026

Open Source Open Markdown Open JSON

Source Summary

Boundary says its team uses **no code reviews**, lets engineers choose their AI tools, and anchors agents with a small **architecture.md** containing durable compiler-layer rules. Agents generate BAML programs, inspect transcripts, surface issues, propose fixes, and compare variants by tool calls, errors, and outcomes.

Practical Implication

The reusable idea is to move assurance into executable invariants and observable agent behavior. Keep architecture context stable, record traces, generate adversarial usage, and let humans triage findings rather than manually reading every generated line.

Agent-Ready Context

Boundary says its team uses **no code reviews**, lets engineers choose their AI tools, and anchors agents with a small **architecture.md** containing durable compiler-layer rules. Agents generate BAML programs, inspect transcripts, surface issues, propose fixes, and compare variants by tool calls, errors, and outcomes.

The reusable idea is to move assurance into executable invariants and observable agent behavior. Keep architecture context stable, record traces, generate adversarial usage, and let humans triage findings rather than manually reading every generated line.

This is one team’s process for a programming language, not evidence that review is broadly obsolete. Agent findings can hallucinate or lack judgment, while claims such as effectively free tracing and deterministic A/B comparisons depend on Boundary’s underlying system.

Connected Context · Feed7 Judgment

Boundary pushes the prior observability-and-validator pattern to an unusually strong conclusion: executable invariants and trace-level outcomes replace routine line review. This confirms that durable architecture rules and behavioral evidence can absorb more assurance work, but conflicts with candidates retaining human code ownership. Its applicability is therefore bounded to systems where behavior is densely observable and enforceable.

Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayerHumanLayer directly challenges Boundary’s no-review practice by arguing that throughput without human code ownership can erode maintainability, identifying the main unresolved tradeoff.From Signal to PR: Anatomy of a Self-Improving Agent — Jason Lopatecki, ArizeArize reinforces the trace-driven investigation loop but retains human review as change scope grows, offering a more conservative boundary for the same observability inputs.How Forward Deployed Engineering is done at Factory — Eno ReyesFactory supports moving assurance into validators and measurable outcomes, while its ceiling for hard-to-measure defects explains why Boundary’s approach depends on unusually observable behavior.The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest & Isaac MillerDSPy’s stable task contracts and evals provide a general form of Boundary’s executable invariants, allowing implementations or variants to change while expected behavior remains fixed.

Context Map

agentcoding#harness-engineering#agent-evals#observability

Uncertainty

This is one team’s process for a programming language, not evidence that review is broadly obsolete. Agent findings can hallucinate or lack judgment, while claims such as effectively free tracing and deterministic A/B comparisons depend on Boundary’s underlying system.