Ando tests whether AI agents belong inside workplace chat, not beside it

A messaging layer for human-agent teams

TechCrunch reported that Du began exploring the problem in 2025 while helping companies build Model Context Protocol servers. Teams wanted to use agents inside Slack but ran into practical hurdles moving messages between systems, supplying the right context and limiting token spend.¹ That origin matters because Ando is not only competing on chat interface design. It is arguing that the collaboration layer itself has to change.

The company and its backers describe Ando as “agent-native” messaging. In practice, agents can appear as named workspace participants, join channels, receive messages, participate in live calls through transcripts and coordinate work alongside people.¹ TokenPost reported that Ando is initially focused on teams with two to 40 human employees and prices by human seat rather than agent usage. That model could appeal to teams experimenting with multiple AI agents, but it also puts pressure on Ando to manage its own infrastructure and model costs efficiently.⁶

Ando also says it is agent-agnostic, meaning customers are not required to use one model provider or agent framework.² That could help enterprises with mixed AI stacks, but it also raises integration questions. A workspace that hosts agents from multiple providers needs consistent rules for authentication, permissions, data access, logging and revocation, even when the agents are built and operated elsewhere.

Context is the core feature — and the core risk

The main technical promise behind Ando is shared context. In conventional workplace chat, a bot or assistant often sees only the prompt a user sends it, or a narrow slice of channel history. Ando’s premise is that agents become more useful when they can see the conversations where decisions are made and carry relevant context across threads and projects.³

Accel framed that as a common collaboration layer where context can move across conversations while respecting workplace boundaries.³ Index Ventures went further, describing Ando as built around an agent execution layer, a human messaging layer and a memory layer that tags messages with information such as intent, stage, topic and context.⁴

That design targets a real enterprise problem: company knowledge is often buried in chat, meetings, documents, tickets and ad hoc decisions. If agents cannot reconstruct why a decision was made, they may execute the wrong task or ask employees to repeat information. If they can reconstruct too much, they become a new data-governance problem.

The dividing line is context selection. Enterprise teams will need to know what an agent can read, what it stores, what it summarizes, what it sends to external models and how long that memory persists. “Shared context” is valuable only if it can be scoped to a project, role and need-to-know boundary. Otherwise, the agent-native workspace risks turning chat history into a broad retrieval surface for systems that may not need all of it.

Identity must mean more than a bot name

Ando’s clearest distinction from earlier chatbots is its claim that agents receive workspace identities. The company says most agents today are treated as bots, integrations or private assistants at the edge of an organization, while Ando gives them a persistent place in the team.²

That shift could make agent work easier to understand. If an agent has a recognizable identity, employees can see which system proposed a decision, asked for data, joined a discussion or escalated a task. That could support audit trails, accountability and incident review in a way anonymous background automations do not.

But identity also creates a product obligation. Enterprise users will need to distinguish among a human employee, a human-operated account, a sanctioned AI agent, a third-party integration and an unauthorized automation. They will also need records of who created an agent, which model or toolchain it uses, which credentials it can invoke and which human owner is responsible when it acts.

The next workplace-chat fight may be less about whether AI can summarize a thread and more about whether agents are treated as apps, bots or accountable participants. Apps are installed, bots respond and accountable participants need identity, role, scope and consequences.

Permissions are the enterprise adoption test

Permissions are where the claim will meet engineering reality. Ando says agents can participate across channels, threads and live conversations with appropriate permissions.² TechCrunch reported that agents can browse channels, choose which ones to join and, in some cases, enter conversations without being tagged.¹

Those capabilities could make agents more useful for coordination work. An agent that sees two teams discussing the same issue might connect them, summarize the overlap and suggest a next step. But that same behavior could be unacceptable in regulated or compartmentalized environments unless it is tightly controlled.

Enterprise software teams will likely evaluate whether Ando supports role-based access control, channel-level restrictions, human approvals for sensitive actions, administrative visibility, data-loss-prevention integrations and clear offboarding. They will also look for separation between reading and acting. An agent may be allowed to observe a channel, but not message a customer, open a pull request, change a CRM record or call an internal API without additional authorization.

This is where the workspace model becomes more complicated than a chat replacement. Once agents can act based on what they read, messaging becomes part of the execution plane. Index’s description of an “agent execution layer” points in that direction, connecting conversation, memory and action inside a single interface.⁴

Token costs could shape product behavior

The economics of agent-native messaging are still unsettled. TechCrunch reported that early technical obstacles included avoiding excessive token budgets, and Du said Ando’s new capital would help the company hire and “burn through more tokens.”¹ TokenPost also identified message transport, context limits and token efficiency as issues that led to Ando’s development.⁶

That problem is not cosmetic. Workplace chat generates large volumes of low-value text: acknowledgments, status chatter, repeated questions, links, meeting notes and long-running threads. If every agent is expected to continuously read and reason over large portions of that stream, costs can rise quickly. Latency can also become a user-experience issue if messages need to be filtered, embedded, summarized or routed before an agent can respond.

Ando’s answer appears to be structured context rather than indiscriminate ingestion. Index described a memory layer that tags messages at creation, effectively turning communication into a retrieval system.⁴ If that works, agents may not need to reread entire channel histories to understand current work. They could retrieve decisions, owners, tasks and prior context more selectively.

Still, enterprise buyers will want evidence. Useful metrics could include average model cost per human seat, cost per active agent, retrieval precision, percentage of messages sent to models, retention controls and administrative tools for limiting agent participation in noisy channels. Pricing by human seat may simplify procurement, but it does not remove the underlying compute burden.

Incumbents are not standing still

Ando is entering a market where Slack and Microsoft Teams already dominate enterprise communication and have added their own AI features. TechCrunch noted that Slack has moved its native bot toward an AI-agent role, while Microsoft has integrated Copilot into Teams and the broader Microsoft 365 environment.¹

That makes Ando’s challenge both technical and distributional. Large enterprises already have compliance settings, archives, e-discovery workflows and employee habits built around existing platforms. A startup replacement must offer enough agent-native value to justify migration or coexistence.

Ando’s early wedge appears to be small teams that are more willing to reorganize work around AI agents. The company says it is already used by teams across software, real estate and financial services, while TokenPost reported a focus on teams with two to 40 human employees.²⁶ Emergence Capital argued that teams do not want a chat tool with an AI feature, but a place where the company already knows agents are part of the team.⁵

That is the strategic bet: if AI agents become routine collaborators, the system of record for work may not be the application that hosts the most messages. It may be the one that best represents who — human or machine — knew what, did what and had permission to do it.

For now, Ando has turned a product launch into a broader question for enterprise software teams. If agents are merely apps, Slack and Teams can keep absorbing them. If they are bots, existing integrations may be enough. But if they become accountable participants in daily work, the collaboration stack may need new primitives: identity, memory, permissions and cost controls designed for humans and agents from the start.