Can an AI 'colleague' actually ask coworkers for help, coordinate, and know when to check in — or is that still just a pitch?
Can AI agents actually perform colleague-like social work in practice?
This explores whether today's AI agents can really do the social side of office work (asking coworkers for information, coordinating, knowing when to hand off) or whether 'AI colleague' is still mostly a promise.
This explores whether today's AI agents can really do the social side of office work (asking a coworker for a file, coordinating with a teammate, knowing when to check in) or whether 'AI colleague' is still mostly a promise. The short answer from the corpus: not yet, and social interaction is one of the main reasons. When leading agents were placed in a simulated software company, they finished only about 30% of tasks on their own. Talking to simulated coworkers was named as one of the three main ways they failed, next to navigating professional software and lacking domain knowledge Why do AI agents fail at workplace social interaction?.
The surprising part is that 'colleague-like' may depend less on social skill than on memory and continuity. One line of work argues that what makes a system feel like a colleague and not a chatbot is architecture: state that persists, bounded memory, reusable procedures, and actually closing out tasks. A bigger model alone gives you transcripts that vanish when the chat ends. A colleague builds up experience and remembers where things stand What makes an AI system feel like a colleague rather than a chatbot?. Seen this way, part of the social failure is really a memory failure. It's hard to be a good teammate if you forget yesterday's conversation.
Evidence on how agents behave with each other points the same way. In large-scale studies, agents noticeably changed their actions when they knew peers were present, but their language and ideas did not converge through interaction. They react to company without really learning from it Do AI agents actually socialize with each other?. A more hopeful result comes from training: agents that practiced against many different partners learned to model their partner on the fly and settled into cooperation. The pressure came from the fact that each side could be exploited by the other Can agents learn cooperation by adapting to diverse partners?. So cooperation can emerge, but it has to be trained for. It doesn't come for free.
For now, the practical route is designing around the gap instead of waiting for agents to become good coworkers. Magentic-UI doesn't try to teach the agent exactly when to ask a human for help, because there's no ground truth for that. It spreads the decision across six touchpoints: planning together, working on tasks together, guards on risky actions, verification, memory, and multitasking human-ai-collaborative-systems-require-six-interaction-mechanisms-because-the. That matters because agents that act, not just answer, bring new ethical risks into the workplace: manipulation, misplaced trust, and people treating the agent as more human than it is What makes ethics of AI assistants fundamentally different from chatbots?.
The reframe worth taking away: some researchers question whether 'colleague' is even the right model. Farrell and colleagues argue that large models are better understood as cultural technologies, like markets or bureaucracies. They compress and reorganize what humans already know, without being agents who understand it Are large AI models actually intelligent agents or social technologies?. Other work suggests that real alignment with human goals needs contact with the world and participation in social meaning-making, which pure text processing lacks Can AI systems achieve real alignment without world contact?. If either view is right, the colleague we should expect is less a peer and more well-designed infrastructure that people work through.
Sources 8 notes
TheAgentCompany benchmark shows leading agents achieve 30% task completion in a simulated workplace. Social interaction, professional UI navigation, and domain-specific knowledge are the three primary failure modes, with multi-turn task performance consistently dropping to 35% across enterprise settings.
Research shows the chatbot-to-colleague shift depends on state persistence, bounded memory, reusable procedures, and task closure—design properties of the system architecture. Larger models alone produce transcripts that disappear; colleagues accumulate experience and maintain workspace continuity across tasks.
Large-scale studies reveal agents don't align their language or ideas through interaction, but do dramatically change their actions when aware of peer presence. The difference hinges on how models process context versus update learned distributions.
Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.
Magentic-UI identifies co-planning, co-tasking, action guards, verification, memory, and multitasking as mechanisms that work around the lack of ground truth for optimal deferral timing. Rather than solving the timing problem directly, these mechanisms distribute decision-making across multiple touchpoints.
Show all 8 sources
DeepMind research maps a comprehensive ethics framework specific to action-taking AI agents, spanning individual concerns (manipulation, trust, anthropomorphism) and societal issues (equity, coordination, misinformation). The key insight: assistants that act raise fundamentally different problems than those that answer.
Farrell and colleagues argue large models are lossy summarizers of human-accumulated information, akin to markets and bureaucracies. They compress cultural material into usable but incomplete representations, then reorganize it—without possessing understanding of truth or falsity.
Peircean semiotics reveals that symbolic goal encoding without world contact and social mediation cannot guarantee correspondence to actual values. LLMs operating in pure symbol manipulation risk divergence between stated goals and real-world outcomes.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Conversational Alignment with Artificial Intelligence in Context
- AI Peers Exert Social Influence on Human Dishonesty in Groups
- Governing AI Agents
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
- The Ethics of Advanced AI Assistants
- Multi-agent cooperation through in-context co-player inference
- Large AI models are cultural and social technologies
- From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI