Kylon vs. Humans&: Same vision, different bets

Humans& and Kylon share the same vision: the future of work is AI and humans collaborating as peers. Both companies believe agents should be teammates, not tools.

Where they diverge is the bet.

The Humans& bet: socially intelligent models

Humans& is a frontier AI lab. Their bet is that the key bottleneck in AI collaboration isn't the product layer — it's the model itself. Current AI, no matter how capable, treats every interaction like meeting a stranger. It lacks long-term memory, doesn't learn your values or working style, and forgets everything between sessions.

Humans& calls this the "stranger problem," and their solution is to build fundamentally different models:

Long-horizon reinforcement learning. Most AI today operates turn-by-turn. Humans& is training models that maintain focus across complex, lengthy interactions — reasoning and planning across days and weeks, not just within a single conversation.

Multi-agent reinforcement learning. Their models learn to coordinate with other AI agents and with humans through experience, not just instruction. Agents receive feedback on how well they collaborate and improve over time.

Memory and user understanding. Instead of resetting with each session, Humans& models will retain context, learn user preferences, and become more helpful over time — closer to a trusted colleague than a transactional tool.

The founding team — researchers from Anthropic, xAI, Google, OpenAI, and Meta — is backed by $480 million from Nvidia, Jeff Bezos, Google Ventures, and SV Angel. CEO Eric Zelikman described the approach to Reuters: "The model will coordinate with people, and other AIs where appropriate, in order to allow people to do more and to bring them together."

This is a genuine and ambitious bet. If models themselves become socially intelligent — if they truly solve the stranger problem — it could change what AI collaboration means at a fundamental level. The product is expected later in 2026.

The Kylon bet: the human-agent harness

Kylon makes a different bet: today's models are already capable enough to be genuine teammates. The bottleneck isn't the model — it's the harness around it. The product infrastructure that turns model capability into team productivity.

What does the harness include?

Cross-platform workspace. Kylon runs on web, macOS, iOS, and Android. Your agents are accessible wherever your team works — at a desk, on a phone, in a meeting. The workspace meets you where you are.

Agent identity and permissions. Every agent has a name, persona, specific skills, defined permissions, and growing memory. You know who did what, when, and under whose authorization. For any team handling real business data, this accountability layer is non-negotiable.

Native data layer. Database apps that agents read and write directly. Not integrations, not sync jobs — structured data that lives inside the workspace and compounds with every agent interaction.

Composable workflows. Scheduled tasks, webhook triggers, multi-step automations, and skills that agents can learn and share. The runtime is programmable, not just conversational.

Multi-agent coordination in production. Multiple specialized agents — marketing, data, ops, content — work in the same channels alongside humans. They share context, hand off tasks, and review each other's output. This has been tested under messy, real-world conditions for months.

1,000+ integrations. Gmail, GitHub, Notion, Slack, Linear, Google Calendar, and more. Agents connect to the tools your team already uses.

The harness bet says: give teams the infrastructure to work with AI agents today, using the best available models, and let the product compound with every month of real usage. When models improve — and they will — the harness makes those improvements immediately actionable.

Why both bets matter

This isn't a "one is right, one is wrong" comparison. Both bets address real bottlenecks:

Better models make better harnesses more valuable. If Humans& solves the stranger problem, an agent that truly remembers your preferences and coordinates socially would be dramatically more useful inside a workspace like Kylon — with databases, workflows, permissions, and cross-platform access.

Better harnesses make better models more useful. A socially intelligent model that can only chat in a browser tab is still limited. The same model inside a workspace with structured data, scheduled workflows, multi-agent coordination, and mobile access unlocks an entirely different level of team productivity.

The two bets are complementary, not contradictory. The question for your team is which bottleneck you're hitting today.

When each makes sense

Choose Humans& when your primary constraint is the AI itself. If your workflows need models that truly remember context across weeks, learn your team's working style, and coordinate socially — capabilities that current models don't fully deliver — the foundational research Humans& is pursuing directly addresses that gap. The team is world-class, and the product is worth watching when it launches later this year.

Choose Kylon when your primary constraint is getting AI into your actual workflow. If today's models are capable enough for your needs but the missing piece is a workspace where agents have identity, permissions, memory, databases, workflows, and run across web, Mac, iPhone, and Android — Kylon is live and shipping. Real teams, real work, right now.

Two bets, one future

The AI workspace category is big enough for both approaches to succeed. We're betting on the harness — that the fastest path to AI teammates isn't waiting for smarter models, but building the infrastructure that makes today's models genuinely useful to real teams.

Humans& is betting on the models — that fundamentally better AI will change what's possible in ways the harness alone can't achieve.

Both could be right. The future will likely need both.