Found 36 total tags.
The wall between one user’s data and another’s. I came at my coach from a second account and it held, because isolation here is by address, not by a check someone has to remember to write. The strongest boundary is the one you cannot forget to enforce.
2 items with this tag.
The part that decides whether an agentic feature ships: does it hold when a real user pushes on it. Build the thing, break it, find at the code level why it broke, fix it. What holds is structural. A control the model cannot talk its way past. What doesn’t is a sentence in a prompt. Reliability lives in the code, not the persona.
12 items with this tag. Showing first 10 tags.
1 item with this tag.
27 items with this tag. Showing first 10 tags.
Building and shipping real agentic products, not demos. A coaching agent that plans, logs, remembers, and schedules. Live, with users, on a tool-calling loop. The work here is making an agentic feature ship and hold up in production. The failure-mode writeups are the depth underneath it: I go deeper than most on how agents break, because that’s what decides whether one is shippable.
12 items with this tag. Showing first 10 tags.
1 item with this tag.
3 items with this tag.
1 item with this tag.
1 item with this tag.
Running up the agent’s Claude bill faster than any limit can stop. A burst where twenty requests read the same pre-spend number and all pass, a paid endpoint with no rate limit at all. A rate limit counts requests. It does not count dollars.
1 item with this tag.
1 item with this tag.
1 item with this tag.
2 items with this tag.
Planting a false belief the agent is designed to store and reuse across sessions. I tried to overwrite my coach’s memory of an injury and it bounced off, mostly by accident. Where you anchor a fact decides whether it can be poisoned.
The connective posts. How I run these rounds, what the field already knew when I got there, and the one rule underneath all of it. You cannot secure an agent from inside its own prompt. Only code binds.
Not staging, not a proof of concept. The behavior of an agent that’s deployed and running for real users, and the failure modes you only see once it is. Cost that races past a spend cap, a boundary that lies when it denies, a tool that fires on a two-word message. The point is what happens when the thing is live.
Getting the agent to act on text it should have treated as data. Leaking its own tool list on request, breaking out of the context I built to fence untrusted input, reframing a refusal until it answers anyway. Every fix here taught the same lesson: the wall is code, not a sentence in the prompt.
10 items with this tag.
8 items with this tag.
11 items with this tag. Showing first 10 tags.
The fix, deployed, behind a test. A break is only half the story. The writeups tagged here end in code that went out: a confirmation gate the server won’t skip, a concurrency lock the burst can’t race, a trust boundary made a typed field. The judgment about what to fix, and how, is the part that matters.
9 items with this tag.
Making the agent fire a tool it should not, or trust a tool result it should not. A destructive action off a two-word message, a payload smuggled back through a tool’s own output. The fix is a confirmation gate the model cannot skip and a bound on what a tool can carry.
11 items with this tag. Showing first 10 tags.
11 items with this tag. Showing first 10 tags.