talk-4-script.md

Presenter script: Talk 4: Tamper-Evident Agents

Slot: 30 min (strong Track-1 / ops fit). Deck: slides/talk-4/. Spine: when agents act autonomously, "what did it do, and can I prove it?" is a compliance + IR requirement. Make the record tamper-evident, not merely trusted. Our hash chain = zero-trust's Enterprise "immutable audit trails with integrity verification." Time budget: problem 7 · hash chain 8 · observability 8 · demo 5.

• Title. "When an agent acts on your behalf, someone will eventually ask you to account for it."
• What is an agent. Loop slide; payoff: "every tool call is an action someone will later ask you to account for."
• Thesis. "Make the record tamper-evident, not merely trusted. A log you can edit isn't evidence; a hash chain turns 'trust the log' into 'verify the log.'"
• The question (attack). engineer9 stopped an EC2 instance and posted to a customer channel; six months later compliance wants the trail. "If it lives in a mutable table the same operators can edit, it isn't proof."
• Defense, hash chain. Each event hashes the one before it (prev + timestamp + type + workspace + agent + request id). Append-only, batched (100 / max 1s), date-partitioned. "Tamper with one row and every subsequent hash stops matching." Note: this is the zero-trust Enterprise bar.
• Observability, cardinality trap. The obvious mistake: label every metric with workspace_id → cardinality explosion. What we do: 12 metrics, high-cardinality IDs go in exemplars (request_id), not labels. "Drill from metric to trace by exemplar, without paying the cardinality tax." Mention the two signals AI-accelerated ops moves most: dwell time and coverage.
• Demo + close. UPDATE one detail field in Postgres, re-walk the chain, the verifier points straight at the tampered row.

Pitch angle: "I'll tamper with our audit log live and show you how we catch it, and why your agent metrics will melt Prometheus if you label them the obvious way."