Research Summary: Humanize: Judgement Engineering for Agentic Coding

Research published on arXiv introduces 'Humanize', a multi-agent orchestration architecture addressing completion reliability issues in agentic coding. The framework operationalises 'judgement engineering' by enforcing strict separation of concerns: a human signs off on planning, a builder agent executes implementation rounds, and a reviewer agent from a distinct vendor determines task completion via deterministic routing and 72 mechanical gates.

Why it matters

As artificial intelligence drastically reduces the marginal cost of code creation, the critical operational constraint shifts to verification and completion assurance. Employing multi-vendor model architectures paired with deterministic gating establishes an essential pattern for mitigating autonomous agent failure modes. This transitions organizational focus from code generation volume to auditable, resilient system governance.

Key insights

• Code-generating agents are ineffective judges of their own completion and output quality.
• The Humanize framework implements judgement engineering using explicit, mechanically enforced boundaries across planning, implementation, review, and learning phases.
• A human approver validates the initial plan contract prior to implementation.
• Reviewer agents are deliberately sourced from an alternate vendor relative to the builder agent to reduce correlated defects across repository states.
• Task routing and 72 mechanical gates are strictly managed by deterministic hooks rather than probabilistic model reasoning.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.08900