Engineering Support Lead

Composio, a San Francisco-based developer of AI agent learning infrastructure, has raised $25 million in a Series A funding round led by Lightspeed Venture Partners. The investment brings the company's total funding to $29 million. Founded in 2023 by CEO Soham Ganatra, Composio addresses learning limitations in AI agents by building shared infrastructure that allows agents to accumulate knowledge and improve through experience over time. The platform provides a shared skill layer that captures and distributes practical knowledge across AI systems. Participating investors included Guillermo Rauch, Dharmesh Shah, Gokul Rajaram, Soham Mazumdar, SV Angel, Blitzscaling Ventures, Operator Partners, Agent Fund, Elevation Capital, and Together Fund. Composio plans to use the capital to accelerate development of its learning infrastructure, which integrates with major frameworks including MCP, LangChain, Vercel AI SDK, and OpenAI Agents.

Introducing Composio Shared Connections. Give every agent on your team access to the apps your company already uses, through a single connection. Most teams have at least one important account that belongs to a single person. It might be the company Gmail account, a Meta Ads account, or a QuickBooks admin login. Everyone needs to work with that account, but nobody wants the password or the API key passed around in Slack. That's why Composio built Composio Shared Connections. You can now connect an account for your team and make it available to everyone in your Composio organization. This enables your team to use their preferred AI, such as Claude CoWork, to analyze emails, draft replies, review Meta Ads performance, or pull QuickBooks insights without sharing account credentials. How to get started. To create a shared connection, log in to your Composio dashboard and go to the Apps page. Find the app you want to connect, open the dropdown next to Connect, and select Connect for your team. You'll sign in again to create a separate connection for your team. If you already have a private connection to the same account, it stays private. Follow the same process and sign in again to create the shared version. Once the shared connection is ready, teammates who don't have their own connection to that app will automatically use the team connection. If someone has connected their own account, their personal connection always takes priority. Invite teammates in a few clicks. After creating a shared connection, you can invite a teammate by entering their email address. You can also invite people later from Settings | Members. If your teammates are already part of your organization, they'll get access to the shared connection. If they're not, they'll receive an invitation to join your organization. Once they join, they'll be able to use all the connections your organization has shared. This means you can bring someone into the tools your team already uses without asking them to connect every account individually. Stay in control. You can manage shared connections from the same place you manage your private ones. Open the app in your dashboard and you'll notice a Connected Accounts section. You'll see your private connections first, followed by the connections you've shared with your team. From there, you can view the usage history for a shared connection, see who used it and when, invite more teammates, or delete the connection. Built for everyday team workflows. Shared Connections make it easier to give teammates access to useful agentic workflows without handing over the account itself. * Share a company Gmail connection so teammates can analyze recent emails or draft replies in Hermes. * Share a Meta Ads connection so the team can review campaigns and uncover performance insights in Claude Code. * Share a QuickBooks connection so your finance team can review transactions and prepare reports without needing the admin login in OpenAI Codex. Your team gets the access it needs. You keep control of the account. Bring secure Shared Connections to the agents you build. If you're building an agent on top of the Composio platform, you can now offer Shared Connections for company-managed accounts, shared mailboxes, and agents that work across multiple users. This is still an experimental feature. Your app starts the connection flow and configures who can use it, and the account owner must still sign in and authorize the connection before it becomes active. An existing private connection is never shared automatically. For more details, see its documentation. Get new posts in your inbox. Subscribe for the latest from the Composio blog.

Claude Code vs OpenCode vs Oh My Pi: speed, cost, success. Advertisements Key takeaways. Is Pi the Best Coding Agent? Pi vs OpenCode vs Claude Code * Claude Code completes tasks in 122 seconds but costs $0.195 per successful task * OpenCode is cheapest at $0.073 per task but slower at 202 seconds * Oh My Pi achieved the highest success rate (17/30) but was slowest at 272 seconds Claude Code is the fastest of four major agent frameworks tested against DeepSeek V4 Flash, completing tasks in 122 seconds on average. It also costs nearly three times more than the cheapest alternative. AI tooling company Composio ran 30 real-world tasks across Claude Code, Codex, OpenCode, and Oh My Pi, using integrations with Gmail, GitHub, Slack, and Notion. The results show that the wrapper around a model matters as much as the model itself. | Framework | Success Rate | Avg. Time/Task | Cost/Successful Task | | Claude Code | ~15-16/30 | 122 seconds | $0.195 | | Oh My Pi | 17/30 | 272 seconds | Not disclosed | | OpenCode | 14/30 | ~202 seconds | $0.073 | | Codex | ~15-16/30 | Not disclosed | Not disclosed | Advertisements What did the benchmark actually test? Composio held the model constant, running DeepSeek V4 Flash through all four frameworks. The 30 tasks involved real-world tool usage: sending emails, interacting with GitHub repositories, posting to Slack, and managing Notion pages. This setup isolates framework behavior from model capability. Seven tasks passed or failed based solely on which framework ran them. That's a 23% variance from the framework alone. The model never changed. The prompts never changed. Only the orchestration layer differed. Why does Claude Code cost so much more? Claude Code clocked $0.195 per successful task despite using the fewest tool calls and generating the least output tokens. OpenCode hit $0.073. That's a 2.7x price gap for comparable success rates. The cost difference likely traces to how each framework structures its context window, manages retries, and batches API calls. Claude Code's speed advantage (122 seconds versus OpenCode's estimated 202 seconds) suggests it parallelizes aggressively. But parallelization inflates token usage if the framework isn't pruning redundant context between calls. For teams running hundreds of agent tasks daily, that 2.7x multiplier compounds. A workflow that costs $73 per thousand successful tasks on OpenCode would cost $195 on Claude Code. Over a month, that's real budget pressure. Oh My Pi won on success rate but lost on time. Oh My Pi completed 17 of 30 tasks, the highest success rate in the test. It also took 272 seconds per task, more than double Claude Code's 122 seconds. For interactive use cases (a developer waiting for code generation, a support agent needing a quick answer), that latency kills the workflow. For batch processing where correctness matters more than speed, Oh My Pi's higher success rate might justify the wait. Composio didn't publish Oh My Pi's cost-per-task, so the full tradeoff picture remains incomplete. Advertisements What should teams optimize for? The answer depends on the job. A startup automating internal workflows with Zapier or Make cares about cost per task. An enterprise shipping customer-facing agents cares about latency. A compliance team cares about success rate. Disclosure. Some links in this post are affiliate links - Logicity earns a commission if you sign up, at no extra cost to you. Logicity only link products Logicity has used or actively recommend. No framework dominated all three metrics. That's the finding that matters. Claude Code is not "the best" agent framework. It's the fastest. OpenCode is not "bad." It's cheap. Oh My Pi is not "slow." It's thorough. The choice is a tradeoff, not a ranking. Logicity's take. This benchmark confirms what many teams suspected: the orchestration layer is a cost center, not a commodity. Teams evaluating agent frameworks should run their own task mix before committing. Composio's test used a specific set of integrations; your Slack-heavy or GitHub-heavy workflow may skew results differently. Also missing: error recovery behavior, which matters when agents fail mid-task. What the test didn't cover. Composio published success/fail, time, and cost. They didn't publish error types, retry behavior, or partial-success rates. A task that fails on the final step after nine correct ones is not the same as a task that fails immediately. Framework error handling matters for production reliability, and this benchmark doesn't measure it. The test also held the model constant at DeepSeek V4 Flash. Different frameworks may perform differently with Claude models, GPT-4, or Gemini. Framework-model pairing effects are real; this benchmark doesn't capture them. Frequently asked questions. Which agent framework is cheapest per task? OpenCode, at $0.073 per successful task in Composio's benchmark. Which agent framework is fastest? Claude Code, at 122 seconds per task on average. Which framework has the highest success rate? Oh My Pi completed 17 of 30 tasks, the highest in the test. Need help implementing this? Evaluating agent frameworks for your stack? Reach out to Logicity's consulting team for a custom benchmark tailored to your integrations and cost constraints. Advertisements Huma Shazia Senior AI & Tech Writer Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in its Editorial Policy.