The Company Teaching AI Agents to Finish the Job

Somewhere, a software company is renaming a field in an API response. There may be no warning, no useful changelog, no ceremony at all. Tomorrow an AI agent will ask that API to do something ordinary - update a customer record, file an issue, send a message - and discover that yesterday's instructions have expired. This is the sort of dull event that tends to ruin a dazzling demo. It is also the event around which Composio has built a company.

The distinction is easy to miss. An AI model can suggest that someone should be added to a Slack channel. Actually adding them requires the right account, the right permission, a valid token, the right API call and some confidence that the action happened. Composio sits in that interval between intention and outcome. It sells access to app actions through SDKs for developers, a hosted MCP connection for existing assistants, a command-line tool and enterprise controls for larger teams. The buyer is paying to spend less time building and tending integrations, not to buy another conversational brain.

The detour that became the destination

Soham Ganatra and Karan Vaidya founded Composio in 2023. Vaidya, the CTO, has described an earlier attempt at coding agents. The models then available did not reliably produce what the team needed. In one enterprise pilot, the supposedly automated effort led to two or three months of manually building and repairing integrations. The failed shortcut made a useful diagnosis: even a much better agent would still need a dependable way into other software.

They made that supporting work the main event. Composio offers tool discovery, authentication, event triggers and execution across apps including Gmail, GitHub, Slack, Notion, HubSpot and Salesforce. A developer can keep the model and agent framework they prefer, then ask Composio for the appropriate tools and a connection to the user's account. An individual can connect an existing assistant through its MCP endpoint. A team can use Shared Connections, introduced in September 2026, to make a company account available without passing a password around. Private connections remain private, and the account owner still approves a new shared connection.

A catalog is a promise with a maintenance bill

Composio's site advertises more than 1,500 app toolkits. That number has the clarity of a supermarket aisle count, and a similar problem: everything on the shelf must still be good. Providers alter fields, permissions, rate limits and whole endpoints. One customer's account permits an action that another administrator has forbidden. A test account is forgiving; a real business account arrives with history, exceptions and old settings.

The company's answer is an internal system called Yang. It uses coding agents to build toolkits, investigate production failures and propose missing actions. Review bots inspect its pull requests, and a human approves every merge. Composio says more than 900 repair pull requests were merged in one recent week. Yang then watches production telemetry for seven days to see whether a fix held. That sequence is more revealing than the headline catalog size: a green test means the code is internally consistent; it does not prove the outside world's API behaves as expected.

There is a lesson here for anyone building agents: put the definition of “done” outside the agent's own account of itself. Yang's sessions can survive a sandbox being stopped and restarted. Third-party error text is treated as untrusted input before it reaches an agent that can write code. Repository permissions enforce the human merge gate. These are ordinary engineering controls applied to a system with unusually energetic workers.

The difference between an answer and an action

Glean offers a useful view from the customer side. It began as enterprise search and moved toward agents that could work inside company systems. Its customers wanted the whole task finished, including writing the result back to those systems. Glean chose Composio for the actions layer rather than build it all itself. The relationship has a personal footnote: Vaidya had been an early engineer at Rubrik, the earlier company led by Glean CEO Arvind Jain.

The smaller stories are less grand and perhaps more useful to a builder. Composio says Opennote added Notion and Google Drive connections in under an hour; Zams used it for sales integrations; 11x used it across enterprise sales workflows. These are company-published case studies, so their quantified outcomes deserve the usual care. The common problem is plain enough: a product team wants its agent to act in software its users already own, while its engineers would rather work on the product itself.

Composio Connect exposes seven meta-tools instead of dumping thousands of app schemas into an agent's context at once. The agent searches for what a task needs, requests authorization when necessary, and runs the chosen action. That design addresses a problem familiar to anyone who has overloaded a model with too many instructions: an enormous menu can make choosing dinner harder.

What the action costs

The business model is unusually legible. As listed in September 2026, the Hobby plan includes 100,000 standard tool calls and 50,000 triggers each month. Pro costs $29 a month with usage credit, then lists $0.0003 per additional standard tool call and $0.003 per additional trigger. Premium third-party services, such as search or browser automation, have separate charges; enterprise terms are negotiated. The free plan is capped. At scale, the relevant unit is not a seat staring at a dashboard but an action performed on someone's behalf.

That meter also clarifies the competitive field. A company can build its own connectors; Nango focuses on OAuth and API connectivity; Arcade sells agent tools and authorization; Merge offers unified APIs in business software categories. Workflow platforms such as Zapier and Workato solve adjacent automation problems. Composio's pitch is breadth plus an agent-friendly execution layer across frameworks, with the upkeep absorbed by its platform. A team whose connector set is tiny and stable may prefer to own it. A team whose compliance rules require a different credential boundary will have to examine Composio's enterprise options closely.

The plumbing gets a vote

In April 2026, the plumbing did vote. Composio reported about 53 minutes of cumulative API degradation and roughly 36 hours when webhook triggers for several apps were unavailable. About 700 customers with active triggers were directly affected. The team first suspected a jump in Slack webhook volume. The real problem was a cleanup job that had stopped pruning a trigger-processing table; as the table grew, the job timed out, and the backlog grew further. A recovery operation caused another disruption. Composio moved trigger processing to a separate database and promised faster incident updates.

That episode gives the company a more credible shape than the usual claim that integrations “just work.” They do work until an assumption under them breaks. In July 2025, Composio announced a $25 million Series A led by Lightspeed, bringing its stated funding to $29 million. The money bought room to pursue an ambitious vision of agents that learn from previous tool use. The public evidence today is more concrete in the present tense: a broad catalog, customer deployments, a repair system and the scars of operating it.

The most copyable idea is almost embarrassingly modest. Pick a recurring action across two apps. Give the agent only the tools it needs, connect accounts with explicit user approval, log the attempted action, and verify the result outside the model's narration. Then arrange for a broken API call to be somebody's measurable problem. Composio would very much like that somebody to be Composio. Whether the market agrees will depend less on how many tools fit on a slide than on whether Tuesday's broken field is working by Wednesday.