|
AGENTS
Agents & AI
Agents draft, route, and propose well. They absorb blame poorly.
I design for that asymmetry before the first irreversible action —
not after the incident bridge.
Autonomy is a feature. Unattended autonomy is a liability.
A careful colleague must still be able to interrupt, explain, and reverse.
|
|
| Failure modes I design for |
- Money move. An agent queues a transfer. Who confirms? Is the confirm logged with a name, not a shared bot token?
- Health tip. A suggestion sounds confident and is wrong. Can a clinician override in one honest click?
- Public statement. Draft goes out under a brand. Who owned the nod before publish?
- Silent loop. Tool calls keep retrying at 2 a.m. Does anyone get paged before harm compounds?
- Eval theatre. A launch checklist passes once. Continuous eval would have caught the drift.
|
|
| The override click |
|
Good agent stacks are not the ones that talk longest. They are the ones where:
- mandate is explicit — what the agent may do without asking;
- high-stakes steps have a named human checkpoint;
- override is one honest click, and the override is logged;
- traces answer “what did it try, and who approved next?” without a scavenger hunt.
Related:
Your Agent Doesn’t Need Your Approval
·
Human Direction, AI Teammates
|
|
|
FIELD NOTE
“A demo that cannot say who owns the next decision is unfinished work.”
That is the bar I use when choosing tools, vendors, and architecture.
|
|
|
|
|
| Special Offer |
|
FREE
Intro call / coffee chat.
No hard sell. Just talk shop
— agents, governance, product.
|
|
|
|
|
|
|
|
|
|
|
|
|
N
Optimized for
Netscape Navigator
|
|
|
Mac
Platinum UI
c. 1996
|
|