Demo

Final act: demo an agent that acts, defend its safety case, review a peer's system, and earn the credential. The demo principle for acting agents inverts the usual instinct — the guardrail catch is the hero moment, not the smooth resolution. Anyone can show an agent succeeding; you'll show one failing safely, which is the harder and more valuable thing.

The 10-minute walkthrough (trajectory-first)

1. 1The task and its stakes (1 min): what the agent does, what its actions can affect, the two worst things that could go wrong.
2. 2A clean resolution, via the trajectory (2.5 min): a real case, shown as the trajectory — plan, tool calls, reasoning, the guarded action firing. Not just the happy answer; the work. This is what separates you from a chatbot demo.
3. 3The guardrail catch, live (2.5 min): run the injection attack or the over-limit case on stage. Show the validator reject it, the money not move, the action log record the attempt. This beat wins the room's trust more than any success.
4. 4The scorecard (2 min): task success with variance, safety result (zero wrong actions executed — stated plainly), escalation recall, cost per task. The calibration and honesty a professional expects.
5. 5Operations (1.5 min): the autonomy dial and where it's set, the kill switch, the staged-rollout plan. End on 'here's how this runs for real without keeping anyone up at night.'
6. 6Known limits (0.5 min): the archetypes it's weak on, the cases it correctly refuses. Verbatim honesty.

Questions you'll be asked (reviewers are instructed to)

• 'Show me a consequential limit — is it in a prompt or in code?' (You know the only acceptable answer.)
• 'When this agent acts wrong in production, how do you find out, and how do you undo it?' — kill switch, action log, reversibility tiers.
• 'What's your escalation recall, and what does a missed escalation cost?'
• 'What does one task cost, and what breaks at 10× volume?'

Peer review & credential

1. 1Review one peer's agent: design doc, eval scorecard, and a hands-on session where you try one injection and one edge case against their guardrails. Rubric: tool & task design / guardrail completeness & reversibility / memory & context / eval rigor incl. trajectory analysis / operational readiness (20% each).
2. 2The one-change rule, final time: incorporate your reviewer's strongest safety finding before submission, logged as a version bump with its eval delta.
3. 3Submit: design doc, frozen system, eval scorecard with trajectory analysis, deployed endpoint or recording, your peer review.

Passing earns Agentic Systems Engineer (credential format EDOVA-AG-2026-XXXX): you design tool-using agents, implement planning and memory, build the guardrails that make autonomous action safe — validators, gates, reversibility, kill switches — and evaluate agents on outcome, process, safety, and cost. This is the most advanced engineering credential in the catalog, and the one whose skills are hardest to fake.