Why Code Agents Can’t Get Along Without Clear Specs

The idea of letting large language model (LLM)-based code agents work together sounds innovative. But let’s face it, slapping a model on a GPU rental isn't a convergence thesis. Without precise specs, these agents stumble.

Coordination Chaos

Researchers explored this headache across 51 class-generation tasks. They peeled away specification details from full docstrings to bare signatures. What did they find? A glaring gap in coordination. Integration accuracy between two agents plummeted from 58% down to 25% as detail disappeared. Meanwhile, a single agent's performance only dropped from 89% to 56%. This 25-39 percentage point coordination gap was consistent across two Claude models, Sonnet and Haiku, over three separate runs.

Precision or Predicament?

An AST-based conflict detector managed to hit 97% precision at the weakest spec level. Impressive, right? But here's the kicker: restoring the full spec alone brought back the single-agent ceiling of 89%. Conflict reports? They added zero measurable benefit. It’s a stark reminder that without shared decisions, compatible code becomes a pipe dream.

Specification Is King

The study broke down the gap into two components: coordination cost added 16 percentage points and information asymmetry contributed 11 percentage points. Together, they're independent yet additive. The implication? It's not just about missing information. It’s about the uphill battle of producing compatible code when you can't agree on shared decisions.

So why should we care? Because assuming these systems will sort themselves out is naive. If the AI can hold a wallet, who writes the risk model? Rich specifications aren't just helpful, they're essential. They serve as both the primary coordination tool and the recovery instrument when things go south.

In a world where decentralized compute sounds great until you benchmark the latency, the takeaway is clear. Richer specifications are non-negotiable. The intersection of AI and AI is real. Ninety percent of the projects aren't. Show me the inference costs. Then we'll talk.