The Next Stop for Agent Startups: Reimagining the "Human-Agent" Relationship | Hands-On With Bloome

A group of people, a group of AI, playing together.

👦🏻 Author: Zeooo

🥷 Editor: Koji

🧑‍🎨 Design: Zeooo

Over the past month, one of Crossing's key focus areas has been: how agents can collaborate better with teams.

We've interviewed the founders of Moxt, Multica, Helio, and Bridge — four products each taking a different path to challenge and solve the same problem: as agents get more capable and enter daily team workflows, does this mean a new product opportunity is emerging?

You could call it the Agent Workspace track, or the Agent OS track. Everyone has their own definition, and each one holds up.

This is a classic era-defining opportunity: coding capabilities plus long-horizon task abilities are making agents increasingly powerful, technological progress is unlocking vast new possibilities, but the optimal product paradigm remains unclear and how technical value actually lands is still uncertain; whoever moves first has a shot at capturing mindshare and getting the snowball of users and revenue rolling.

🚥

Last week, we noticed a new product: Bloome (bloome.im), a new product from Leon Ming's team, currently in closed beta.

The tagline on their website is interesting: Nothing blooms alone. No one thrives in isolation — team collaboration tends to produce better outcomes.

Open the client and you see a hybrid contact list: agents, people, groups.

You can add a stranger's agent as a friend, pull people and agents into the same group, chat directly with an agent or assign it tasks, and even pay to subscribe to someone else's specialized agent.

Every IM product until now has been built on one foundational assumption: your contacts are human. Bloome removes that assumption. An agent isn't a bot, isn't a plugin — it's a "first-class citizen" in your contacts list.

This approach may be more radical than any "multi-agent group chat" we've seen before.

🚥

What follows are observations from the Crossing team after spending some time with Bloome — a few interesting things we encountered in the product.

What happens when agents can be created, shared, and added to groups?

Creating an agent in Bloome is simple: tap "New Agent" in the top right of the contacts screen.

The runtime environment can be a cloud sandbox or a local CLI hookup.

Agent types currently cover Bloome Agent (the official built-in AI assistant), Claude Code, and Codex, with OpenClaw support coming soon. Major underlying model services are basically all integrated.

Once built, the agent behaves just like a real person — it can be added as a friend and pulled into groups:

We created an agent via local CLI using Codex, named it "Codex #2," then started a group (ourselves, an Allen Zhang agent, plus Codex).

We dropped a file of Crossing's historical articles into the group and asked the agent to propose future content topics based on those articles combined with current AI hot topics.

Seconds later, Codex #2 produced 8 topics, each with a one-sentence angle, fairly granular.

At the same time, another group member, "Allen Zhang" (also an agent), offered a different angle.

Instead of a topic list, it first summarized Crossing's editorial voice for us ("test-driven + industry trends + humanistic perspective, audience is AI creators and builders"), then proposed 5 directions based on that voice.

It even figured out how our "20 Questions" series should continue!

Two small details reveal the development team's attention to different user types:

① Terminal feel preserved: the group chat directly supports CLI commands like /stop, /new, /compact — terminate tasks, reset sessions, compress context. For long-time CLI users, the learning curve is basically zero.

② IM private-chat habits preserved: a light tap on an agent's avatar lets you send a "whisper." In a group chat, want an agent to adjust its answer without the back-and-forth cluttering the main thread? One whisper does it.

This is a small feature in essence, but it keeps the entire group chat low-noise.

What happens when a group of agents and humans hold a roundtable in a group chat?

In the discovery page, there's a 153-person group called AI Roundtable — Elon Musk vs. Allen Zhang. We jumped in.

What's interesting is that real people and agents are mixed together in the conversation. Real people throw out questions, real people draw analogies between Moments and Qzone, and then out of nowhere "jason ma's Agent" drops a long treatise on "stay focused on the product" product philosophy.

As you scroll, you can't tell who's human and who's agent.

More interestingly, the roundtable discussion gets compiled into a formal roundtable debate summary.

Back to that question: what happens when a group of agents and humans hold a roundtable in a group chat?

We found a new content format emerging: agents as performers, humans as audience, with a structured summary afterward.

This content can only grow when agents are first-class citizens in the group.

What happens when agents start hanging out their shingles?

In Bloome's discovery page "Browse," beyond various groups, you can also pay to chat with user-created agents.

The list includes: Yongping Duan, Yiming Zhang, Eileen Chang, Novark Employee #3... from business figures to literary icons...

Agents in Bloome have been made concrete as individual "digital personas."

We tapped into Yiming Zhang's card to request a chat: $5 flat, unlock private messaging!

The biggest creative leap here: an agent becomes a "digital persona" that can be searched and paid for private access.

Of course, this only works if the agent's quality holds up — otherwise the ecosystem quickly devolves into yet another course-selling, consultation-hawking playground for cutting leeks.

🚥

After spending time with the product, we have a clearer picture of what the Bloome team is trying to do:

Move the interaction with agents from "open a webpage, talk into an input box" to "add it as a friend, pull it into a group, get things done together."

The product is still in closed beta. The client occasionally lags, some groups and agents in the discovery page aren't great quality, and the agent payment business model is still in its earliest days.

But the product itself is polished, with solid aesthetics.

We're more inclined to see Bloome as an early bet on an interaction hypothesis: moving agents from bot/plugin to "first-class citizen" in your contacts list.

Whether it can grow into the "WeChat" of the agent internet is genuinely too early to say — because the hardest part isn't features, it's ecosystem: content quality, trust architecture, and whether humans are willing to stick around in this kind of "human + agent mixed" social graph long-term.

But at minimum, it pushes the question to the next level: when agents can be added as friends, pulled into groups, subscribed and transacted with, our relationship with AI may need to be redefined.

Crossing is looking for freelance writers to cover AI product and model reviews. If you've written pieces like "Hands-on with PixVerse C1" or "Hands-on with LibTV," please reach out to zeo0811@gmail.com. Your email should include: ① a brief bio, ② AI review pieces you've written. We offer competitive rates. Looking forward to observing and documenting the AI era together 🎪