Already 7 months old when the scout caught it.
Agent behavioral contracts
What it is
Agent Behavioral Contracts (ABC) apply the software-engineering idea of design-by-contract to autonomous AI agents. Instead of relying only on a natural-language prompt, an agent gets a formal contract made of four runtime-enforceable parts: preconditions (what must be true to start), invariants (what must stay true throughout the session), governance policies (constraints on which actions are allowed), and recovery mechanisms (what happens when a soft violation occurs).
Why it's emerging now
As agents move from single tool calls into long-running, multi-step, sometimes multi-agent sessions, the industry's existing safety vocabulary — guardrails, evals — mostly checks a single output after the fact rather than bounding behavior continuously. ABC-style frameworks instead try to hold an agent to a specification across a whole session, including a probabilistic compliance measure that accounts for LLM non-determinism. The term has surfaced independently from an academic paper (an Accenture researcher, on arXiv) and from a practitioner's blog series and open-source runtime library that predates it — two unconnected efforts converging on the same name for the same idea.
How it relates to what you know
It sits next to guardrails and evals (both known) but works differently: guardrails typically filter or block a single input/output, and evals score behavior after the fact, while behavioral contracts are enforced continuously across the lifetime of an agent session, with defined recovery paths rather than a simple block/allow.