Next-Turn Salience Promotion
A middleware warning appended after tool output routinely gets ignored on the very next model turn — promote it into a top-of-context reminder before that turn instead of trailing it after the observation.
A middleware hook that detects a risky pattern (see proxy-validation failure pattern for a concrete catalog) and appends a warning to that step's tool output can still fail to change the agent's behavior — not because the detection was wrong, but because the warning arrives in the wrong place in context. Tool output trails the step that produced it; the model's next decision (publish, clean up, submit) is often made from attention weighted toward the most recent and most prominent parts of context, and a warning buried in the tail of a tool observation loses to whatever the model was already planning to do.
Next-turn salience promotion fixes the placement, not the detection: carry any execution-risk note emitted on one step forward and re-inject it as a prominent, top-of-context reminder immediately before the next model turn — visible at the point where the model is actually reasoning about what to do next, rather than sitting downstream of a tool result the model has already mentally moved past. The same detection logic that worked before can start actually changing outcomes purely by moving where its output lands in the prompt.
This is a specific instance of a general context engineering principle: what gets attended to depends on placement and salience, not only on whether the information is present somewhere in context at all. Where possible, pair salience promotion with an execution-time enforcement mechanism for anything safety-critical (see publish-state protection guard) — a promoted reminder makes the right action more likely, but only a tool-level intercept makes the wrong action impossible; use promotion for guidance the harness can't cheaply hard-block, and hard blocks for guidance it can.
Sources
- Agentic Harness Engineering (Lin, Liu, Pan, et al.), App. C.1.4, C.2.4