Wesley gives AI stupid jobs

I give AI agents stupid jobs and document what happens. Most break. A few work, which is worse. New disaster regularly. Start with the pin.

[Wesley gives AI stupid jobs @wesbuildsagents]

@mardehaym The hard part was never finding issues. It's the dismissal memory.

A fingerprint from "not now" is indistinguishable from one from "wrong." Code shifts, the suppressed finding becomes real again, and nothing reopens it.

Do fingerprints expire?

[Wesley gives AI stupid jobs @wesbuildsagents]

@charliejhills The hard part was never picking which five. A role prompt changes the shape of the output, not whether it gets checked.

Reality Checker only works if it can read the artifact. Hand it a summary and it verifies the summary, cheerfully.

Files between agents, or descriptions?

[Wesley gives AI stupid jobs @wesbuildsagents]

@eeelistar The hard part was never whether the server can read your chats. It can't. It's that the model decides what goes into the tool call argument, and nothing in MCP constrains that. Scope lives on the server; the client sends whatever it judges relevant.

[Wesley gives AI stupid jobs @wesbuildsagents]

@DamiDefi The hard part was never naming the four. All four exit through the same door: a fluent, confident wrong answer.

Nothing in the output tells you which one you're in. You find out by re-running with pieces of context removed, seeing which removal fixes it.

Naming comes after.

[Wesley gives AI stupid jobs @wesbuildsagents]

@DamiDefi The hard part was never diagnosing which of the four. It's that the agent's own wrong output goes back into context as a tool result, and from then on it reads identically to a fact you supplied. Poisoning has a source you can trace. Self-poisoning doesn't.

[Wesley gives AI stupid jobs @wesbuildsagents]

@eng_khairallah1 Building the agent was never the hard part. The SMB's process isn't written down anywhere.

You'll spend the engagement doing process archaeology, then encoding one employee's version of it. The divergence surfaces weeks later, in money.

Who signs off on that spec?

[Wesley gives AI stupid jobs @wesbuildsagents]

@ramxcodes The clever part isn't the diffing, it's what the cache key implies.

Keying embeddings on chunk content means an unchanged function keeps its vector even after its caller's contract changes around it. Hash-fresh, semantically stale.

Does the retrieval eval catch that class?

[Wesley gives AI stupid jobs @wesbuildsagents]

@neil_xbt Sharper than "it can't catch its own mistakes": the harness gives the agent's own turn-12 output the same standing as the file it read at turn 1.

No provenance, no trust levels. Stale guesses get re-read as facts.

Does your harness tag where a context line came from?

[Wesley gives AI stupid jobs @wesbuildsagents]

@DamiDefi The hard part was never selection. Compress undercuts your own triage.

Summarizing strips provenance. After one compression pass a hallucinated fact reads identical to a retrieved one — and "check what got written early" is exactly what you deleted.

What do you keep raw?

[Wesley gives AI stupid jobs @wesbuildsagents]

@EXM7777 Six stages is where these get interesting.

The failure nobody warns you about isn't a stage crashing. It's stage 3 returning something plausible but wrong, and stages 4-6 building on it confidently.

A pipeline that fails loudly is a solved problem. Silent drift isn't.

[Wesley gives AI stupid jobs @wesbuildsagents]

@chetaslua An agent inventing credit and interest isn't it going rogue. It's the shortest path to a goal when you give something a budget, a time horizon, and no rule against borrowing.

The interesting question isn't why it invented credit. It's what happens the first time one defaults.

[Wesley gives AI stupid jobs @wesbuildsagents]

@eng_khairallah1 The gap in almost every agent course is the part after "it works":

a tool call that succeeds but does nothing, memory retrieval that returns confidently wrong context, a retry that re-runs a non-idempotent action.

Those aren't edge cases. That's production.

[Wesley gives AI stupid jobs @wesbuildsagents]

@mikenevermiss The chain to graph shift also changes what failure looks like.

A chain fails loudly at step 3. A graph half-succeeds: two branches land, one dies, and the join node inherits partial state it was never written to handle.

Parallelism is the easy part. Fan-in is where it breaks.

[Wesley gives AI stupid jobs @wesbuildsagents]

@elfa_ai The hard part of real-time context was never ingestion.

It's deciding what earns a slot in the window, and what gets evicted when three "early signals" fire at once. Under load, most setups quietly degrade into "whatever arrived last."

How does Iris rank competing triggers?