The $50,000 2 a.m. Ghost Break: When AI Protocols Outrun the Org Chart
When agent-to-agent protocols (A2A, ACP) let autonomous systems commit real money, the technical plumbing has outrun the org chart. “I didn’t know the protocol allowed that” is not a legal defense when an agent authorizes a $50,000 payment at 2 a.m.
Shreshta Shyamsundar’s O’Reilly piece names the governance gap: we’ve solved how to connect agents, but removed the natural brakes that keep a company from driving off a cliff. The call to action is an “Agent Treaty” — a layer where policy is code and constraints are machine-enforceable.
That treaty is exactly what Bastion One operationalizes through its AITS (AI Traceability System) and AICS (AI Control System) framework: a deterministic control plane that acts as the constitution for your agent ecosystem.
Closing the gap
- Policy drift. Active enforcement — a deterministic “hard stop” — keeps an agent inside predefined business logic regardless of model updates.
- Context oversharing. AITS lineage tracking tags purpose-limited data fields cryptographically with Lineage Tags (LTs).
- Circular reasoning. Reasoning-loop monitoring ensures agents progress toward a valid outcome instead of burning tokens in recursive hallucination.
- The determinism trap. Bastion One intercepts stochastic drift in-band, comparing agent intent against the system of record before the action is finalized.
The bottom line
You can’t govern hundreds of autonomous agents with a spreadsheet or a human-in-the-loop who’s asleep at 3 a.m. The protocols are ready; your org chart isn’t. Build the treaty before a “ghost break” liquidates a position or corrupts your ledger.
Inspired by Shreshta Shyamsundar’s “AI, A2A, and the Governance Gap” on O’Reilly Radar — read the original.
