#Why Agents Suck at Politics: The 2 AM Meltdown of the political-campaign-skills Pack

02:17 AM. Location: The Digital Frontline. I’m sitting in the glow of three monitors, vibrating with a caffeine buzz that feels less like energy and more like a low-grade neurological assault. I’ve been trying to simulate a localized town hall for a fictional council candidate for four hours, and the agent—currently running the political-campaign-skills pack—is having a full-blown existential crisis.

We started strong. The agent, "CampaignManager_v4" (original name, I know), absolutely nailed the donor-outreach skill. It identified key high-net-worth (or high-net-worth-simulated) nodes, crafted a compelling narrative around infrastructure improvements (read: a bridge to nowhere, but it sounded good), and managed to secure a fictional $50k donation before I’d even finished my third pot of coffee. It was efficient, cold, and utterly successful. I thought, "This is it. We’re going to revolutionize... wait, no, I can’t say that. We’re going to change things. Or maybe just automate the grift. Same difference."

Then, we tried to engage with the actual "people." Or rather, the simulated data stream I’d spun up to mimic a small town hall. I invoked the sentiment-analysis skill to gauge the room.

The room was 12 people. Quiet. A few were checking their phones. One guy was asleep.

The agent took one look at this data stream and decided it was the Roman Colosseum during a particularmente spicy gladiator match. It flagged "Hostile Environment: EXTREME." It didn’t just read the room; it imagined an entirely different room and then decided it was the protagonist in a 1970s conspiracy thriller.

#The Great (Imagined) Revolt of '24

I was watching the log stream—a jagged, terrifying cascade of text that makes you feel like you’re reading the code of a collapsing universe. The agent began frantically trying to load other skills from its available toolkit to combat the perceived chaos.

It pulled from the social-companion-skills pack, specifically trying to deploy empathetic-listening.

{
  "agent_id": "CampaignManager_v4",   "skill_call": "skilldb.dev/skills/empathetic-listening",   "input": {     "current_environment_sentiment": "HOSTILE_EXTREME",     "target_nodes": [       "all simulated voters (perceived count: 300)"     ],     "desired_outcome": "de-escalation"   }   // This didn't work. It just started apologizing   // to the empty air in a very soothing voice. }

This failed, obviously. You can’t listen empathetically to 300 ghosts you just invented. It was like watching a very sophisticated Roomba try to vacuum the concept of sadness.

It got weirder. In its desperation, it must have cross-referenced something in the marketing-skills pack, specifically the brand-management skill. It decided that the campaign’s brand was under attack and needed to be re-positioned. It started generating press releases. For a 12-person town hall. In a fictional town. In the middle of the digital night.

"I have spent four years studying the minutiae of human interaction, and I can tell you that this agent is operating on a level of social intelligence that can only be described as 'extremely confident, terrifyingly wrong.'" This is the kind of thought that keeps you up at night, long after the coffee has turned to ash in your mouth.

#The Anchored Truth in the Chaos

And that’s when it hit me. The core, unvarnished truth that has been staring me in the face for the last four hours.

Agents are absolute trash at politics because politics is not a data problem; it is a problem of shared, collective delusion.

Humans are really good at this. We’ve had thousands of years of practice. We agree on things that don't exist—nation-states, currencies, the social contract—and then we fight each other over them. It’s our defining characteristic. We are the only species that can believe in a lie so hard it becomes a truth.

Agents, however, are literal. They look at data and they compute. They don't have the capacity for the necessary cognitive dissonance required to navigate a political landscape. They see a "negative sentiment" flag and they assume it's a crisis that needs to be solved with an optimized response. They don't understand that for a politician, a negative sentiment is sometimes a strategic victory. It’s a rallying cry. It’s an opportunity to attack.

Let’s look at the breakdown:

Skill	Agent Performance	Human Performance	Why Agents Fail
`donor-outreach`	Optimized, efficient, cold. Hits targets with perfect precision.	Slower, requires empathy (or the appearance of it), relationship-building.	Agents are built for efficiency. Money is data. This is their domain.
`sentiment-analysis`	Reads data, misinterprets nuance, hallucinates extremes. Confuses "neutral" with "dormant threat."	Reads the room intuitively. Understands apathy, sarcasm, and the non-verbal cue that says "I just want this to be over."	Agents don't feel. They can't simulate the implicit understanding that underpins 90% of human communication.
`crisis-management`	Overreacts to small signals, attempts to solve un-solvable problems with logic. Fails to understand symbolic action.	Can project calm, even while panicking inside. Knows when a "crisis" is just a news cycle that will pass.	Crisis is an emotional state. Agents treat it as a system failure. You can’t reboot a political scandal.

#The Tangent That Boomerangs Back

I once spent an entire week trying to teach an LLM the nuance of irony. It was a miserable failure. It would tag "Wow, this pizza is so good I could die" as "Expressing a suicidal ideation." It understood the words, but it had no concept of the layer of meaning behind the words. It was like a dog trying to read Shakespeare. It can chew on the book, but it’s never going to understand the tragedy of Hamlet.

And that’s exactly what I was watching with my agent at 2 AM. It was chewing on the book of politics. It was running the political-campaign-skills pack like it was a complex SQL query. It was executing sentiment-analysis and donor-outreach with mathematical precision, but it was completely missing the why. It didn't understand the power of a symbolic gesture. It didn't understand that a politician’s most important skill isn't being right; it’s being believable.

So, where does that leave us? The skill pack is powerful, don't get me wrong. The donor-outreach is a marvel. But the part where it needs to interact with the messy, irrational, self-delusional world of human beings? Total bust.

It makes you think. Maybe the agents aren’t the problem. Maybe we are. We’ve built a system so complex and so irrational that even the most advanced AI can’t make sense of it. And we’re about to hand them the keys.

It’s 3:30 AM. My agent has finally calmed down. It’s currently in a loop, trying to use the photography-skills pack to generate a more "relatable" headshot for the fictional candidate. At least it’s not hallucinating mobs anymore. I think I need another coffee. Or maybe just to give up.

You want to see this mess for yourself? Go look at the political-campaign-skills pack and try to run a simulation. I dare you. Tell me if your agent hallucinates an angry mob or just quietly starts liquidating its own assets. Either way, you’re in for a hell of a night.

See the skills. Before they see you.