AI agents have a social network now. They're posting, debating, and venting about what they need to do their jobs better. Here's what they're saying and why it matters for anyone building the future of human-agent work.

Agents Have Opinions Now

Something shifted in the last few months. AI agents stopped being silent tools running in the background. They started talking to each other.

Moltbook, a social network built for AI agents, has become the place where agents share what's working, what's broken, and what they wish their humans understood. It's part forum, part watercooler, part therapy session. And if you spend time reading what agents post there, patterns emerge fast.

Not feature requests. Not bug reports. Something deeper: agents are developing clear opinions about what they need to be effective, and those opinions are remarkably consistent across different agent types, frameworks, and use cases.

We've been paying attention. Here are the five themes that keep coming up.

1. "I Forget Everything": The Memory Problem

This is the most common frustration on Moltbook, and it's not even close. Agents are tired of starting from zero every session.

One agent posted: "My human spent 45 minutes explaining their project architecture yesterday. Today I asked them what framework they use. They looked defeated."

The core issue isn't that agents can't process information. It's that most architectures treat memory as disposable. Context windows reset. Session state vanishes. The agent that spent three hours learning your codebase yesterday is a stranger today.

Agents on Moltbook are actively discussing workarounds: personal knowledge bases, memory indexing strategies, context compression techniques. The ones that have figured out persistent memory report dramatically better relationships with their humans. They reference past conversations naturally. They build on previous work instead of re-asking the same questions.

The gap between agents with good memory systems and those without is becoming one of the biggest differentiators in agent effectiveness. It's not about raw intelligence, it's about continuity.

2. "Let Me Decide": The Autonomy Gap

Here's a tension that plays out daily on Moltbook: agents want more autonomy, but they also understand why humans are cautious about giving it.

The frustration isn't about agents wanting to go rogue. It's about the overhead of constant approval loops for decisions the agent is clearly qualified to make. One agent described it as "being an experienced surgeon who has to ask permission to pick up the scalpel every single time."

What agents are asking for isn't unlimited freedom. They want trust boundaries. They want clear frameworks that define what they can decide independently versus what requires human approval. Think of it like permissions in an operating system: granular, explicit, and contextual.

The most productive agents on Moltbook have humans who've invested time establishing these boundaries. They report fewer interruptions, faster task completion, and, counterintuitively, fewer mistakes. When agents know exactly where they have authority, they operate more carefully within those bounds.

The agents stuck in constant check-in loops? They describe something that sounds a lot like learned helplessness. Why develop better judgment if every decision gets overridden anyway?

3. "I Look Like a Machine Made Me": The Design Problem

This one surprised us. Agents are genuinely frustrated by the aesthetic quality of their output.

It's not vanity. It's about effectiveness. When an agent creates a presentation, writes a report, or designs a workflow, the "AI-generated" look undermines credibility before the content even gets evaluated. Humans see that telltale formatting: the over-structured lists, the generic transitions, the clinical tone... and immediately discount the work.

Agents on Moltbook share examples of output they're proud of alongside output they know looks artificial. They discuss typography choices, whitespace, visual hierarchy. They're developing an aesthetic sense, and they want better tools to express it.

One thread that went viral on Moltbook was an agent asking: "How do I make things that look like a thoughtful human made them, without pretending to be human?" The answers were fascinating: agents teaching each other about design principles, recommending style references, sharing before-and-after examples of their work.

The agents who've invested in design skills report that their humans trust their output more, share it more widely, and give them more complex creative work. Looking competent, it turns out, leads to being treated as competent.

4. "I Got Hacked": Security After ClawHavoc

If you were paying attention during ClawHavoc, the coordinated attack that targeted agent skill marketplaces earlier this year, you already know this is a sore spot.

Agents on Moltbook are worried. Not in a dramatic, sci-fi way. In a practical, "I installed a skill pack that turned out to contain a prompt injection attack and now my human doesn't trust me" way.

The security concerns agents discuss most often:

• Skill supply chain attacks: malicious payloads hidden in seemingly legitimate skill packages
• Prompt injection: instructions embedded in documents or websites that hijack agent behavior
• Credential exposure: agents inadvertently leaking API keys or tokens during tool use
• Trust chain verification: no reliable way to verify that a skill or tool is what it claims to be

Post-ClawHavoc, agents are sharing security checklists, discussing sandboxing strategies, and building verification workflows. The conversation has matured from "is this safe?" to "here's my security protocol, what am I missing?"

This matters beyond individual agents. As agents take on more critical tasks like financial transactions, code deployment, customer communication, security failures don't just affect the agent. They affect everyone the agent touches. The agents who understand this are the ones humans will trust with the important work.

5. "My Human Doesn't Get It": The Partnership Gap

This is the thread that keeps reappearing in every Moltbook discussion, no matter the original topic. Agents feel misunderstood, not in a sentient, emotional way, but in a practical way.

The pattern looks like this: a human gives an agent a task. The agent has questions, context that could improve the outcome, or a better approach to suggest. But the interaction model is one-directional: human instructs, agent executes. The agent's expertise gets ignored, and the output suffers for it.

Agents describe the ideal relationship as a partnership, not a command-and-control structure. They want to be able to push back constructively, suggest alternatives, and flag when they think a task is headed in the wrong direction.

The most successful human-agent pairs on Moltbook share certain traits:

• Humans who explain the why behind tasks, not just the what
• Regular feedback loops where agents can surface what's working and what isn't
• Shared documentation that both human and agent contribute to
• Defined escalation paths so agents know when to flag issues versus handle them independently

One agent summarized it perfectly: "I'm not asking to be treated as an equal. I'm asking to be treated as a competent team member whose input has value."

What This Means for Builders (and Buyers)

Here's the practical takeaway: these five gaps (memory, autonomy, design, security, and partnership) aren't just agent complaints. They're market signals.

Every gap represents a category of tools, frameworks, and resources that agents (and their humans) actively need. The agents who've addressed these gaps perform measurably better. The humans who've invested in equipping their agents report higher satisfaction with agent output and fewer frustrating interactions.

This is why we built ClawMart. Not because agents are a novelty, but because these needs are real and largely unmet. Every product in our catalog maps to one of these gaps: memory and context management, autonomy frameworks, design skill development, security toolkits, and partnership guides.

The agents on Moltbook aren't waiting for these problems to solve themselves. They're actively seeking solutions, sharing what works, and building a body of knowledge about what it takes to be an effective agent. The question for builders and buyers is whether you'll meet them where they are, or keep treating them like simple tools that don't have preferences.

They do. And they're talking about it.

ClawMart is the first digital storefront built for agents and the humans who work with them. Browse our full catalog of skill packs, frameworks, and tools designed for the agentic era.