Moorcheh.ai on X: "Every one I've watched fail did the same thing: perfect demo, great for three weeks, quietly useless by month three. Nothing broke. They just got full. Wrote up what's actually happening ↓ https://t.co/tB4pIepgIQ" / X<br>Post
Log inSign up
Post
Moorcheh.ai
@moorcheh_ai
Every one I've watched fail did the same thing: perfect demo, great for three weeks, quietly useless by month three. Nothing broke. They just got full. Wrote up what's actually happening ↓<br>Nothing Broke. Your Agents Just Got Full.<br>The demo was honest. Month three is also honest. Between them sits a failure that isn't the model's fault; and it has killed more agentic projects than any capability gap I've seen.<br>I have never lost an agent deployment to a model that wasn't smart enough.<br>Not once. Every deployment I've watched go sideways went sideways for a different reason, and it's always the same reason.<br>If you run agents in production, you already know the shape of this. Let me describe it and you tell me how close I get.<br>The demo is not lying to you<br>The POC is beautiful. It genuinely is.<br>You wire up three or four agents, you point them at a real workflow, you run it in front of the customer, and it works. Not "works with caveats." Works. The customer's ops lead goes quiet in that specific way people go quiet when something they didn't believe in just happened in front of them.<br>You ship it. Weeks one through three are great. People start routing real work through it.<br>Then somewhere around week six, someone on the customer's side says something small. "It used to be faster at this." Or: "It gave Priya a different answer than it gave me."<br>By month three you are in a standing weekly call where an engineer on your team walks through what the agents did wrong that week, and quietly adjusts something. Prompt tweaks. A new guardrail. A rule that says "don't do the thing you did on Tuesday."<br>The system is not broken. Nothing throws. Nothing pages. It just isn't good anymore, and it takes a human babysitting it to stay even mediocre.<br>That's the failure. It's not a crash, it's a slope.<br>Three symptoms, one cause<br>Every time I've traced one of these, the decay showed up in three places at once. They look like three separate problems. They aren't.<br>One: the fleet stops agreeing with itself.<br>Two agents, same company, same customer, different answers. The support agent thinks the account is on the old plan. The billing agent knows about the upgrade. Neither is wrong given what it knows. Nobody owns the difference.<br>This is the best-documented failure in the field and almost nobody budgets for it. The MAST taxonomy out of Berkeley. 14 failure modes, built from analysis across seven popular multi-agent frameworks with six expert annotators and a Cohen's kappa of 0.88, puts roughly 37% of observed multi-agent failures in a single bucket they call inter-agent misalignment.<br>Not model error. Not bad reasoning. Agents talking past each other, losing state at handoff, producing conflicting outputs. Over a third of everything that went wrong, in the one category no single agent is responsible for.<br>Two: instructions pile up and start fighting.<br>The user said "always use Postgres" in March. In June they said "we're moving to Snowflake for anything analytical." Both are in the history. Both are true statements about different moments. The agent has no notion of "different moments" it has a similarity score.<br>So the agent retrieves whichever one embeds closer to the current query, and its behaviour becomes a coin flip that looks like a decision.<br>Microsoft Research published the cleanest measurement of this I've seen: take fully specified tasks, break them into pieces, reveal one piece per turn the way a real user actually does, and every top model tested (open and closed weight) drops an average of 39% against the single-turn version. Over 200,000 simulated conversations.<br>Here's the part that should worry anyone running agents. That drop decomposes into a small loss of raw aptitude and a roughly 112% increase in unreliability. The models didn't get dumber. They got inconsistent.<br>And the failure is sticky. Their line for it: when models take a wrong turn in a conversation, they get lost and don't recover. Early wrong assumption, no self-correction, and every downstream step inherits it.<br>That is exactly what a fleet running unsupervised for a quarter looks like from the outside.<br>Three: focus dissolves under volume.<br>This is the one teams argue with me about, because the intuition is so strong that a bigger context window fixes it.<br>It doesn't. Chroma tested 18 frontier models (GPT-4.1, the Claude 4 family, Gemini 2.5, Qwen3 ) on how performance holds as input length grows. Every single one degraded. Not most. All of them.<br>And the degradation starts well before the window is anywhere near full. A 200K window can show meaningful decay at 50K. The decline is continuous, not a cliff, which is why nobody notices it happening.<br>Anthropic's own engineering write-up...