Composio

At Composio, we are building infrastructure that allows agents to communicate with the tools you use for work including Github, Gmail, Notion, Salesforce, etc. We are a small team of engineers wrangling problems from context to search, that help us provide the most capable bridge between your agents and your tools. We raised a $25M Series A from Lightspeed https://lsvp.com/stories/investing-in-composio-building-the-backbone-of-ai-agent-intelligence/ with some incredible angels like Guillermo Rauch (CEO of Vercel), Dharmesh Shah (CTO of Hubspot), Gokul Rajaram. Beginning of this year we 3x our ARR, our customer range from your friends in the YC batch to Wabi, Glean, Zoom and many more. WHAT YOU'LL DO? - build large evals with real tool calling data, measuring where models suck in long horizon tool execution - work on search problems for finding semantically similar tools and cached tool execution paths and plans. - train large agentic harness systems to improve session accuracy with millions of real tool calls as baseline data - SFT on our agentic traces and RL models on top of our agentic harness and app sandboxes. "MUST HAVES" if you are very good, nothing is a must per-se - research - you can independently execute against the research goals - you can prototype and test experiments very quickly - you can work with product and engineering teams to take research ideas to production in a matter of days - typist — you can write docs well and explain complex ideas clearly - human — you build trust and admit what you don’t know

At Composio, we are building infrastructure that allows agents to communicate with the tools you use for work including Github, Gmail, Notion, Salesforce, etc. We are a small team of engineers wrangling problems from context to search, that help us provide the most capable bridge between your agents and your tools. We raised a $25M Series A from Lightspeed https://lsvp.com/stories/investing-in-composio-building-the-backbone-of-ai-agent-intelligence/ with angels like Guillermo Rauch (CEO of Vercel), Dharmesh Shah (CTO of Hubspot), and Gokul Rajaram. We've 3x'd our ARR this year, and our customers range from your friends in the YC batch to Wabi, Glean, Zoom and many more. the action layer for AI agents is the least-solved part of the stack. this role works at that frontier: embedded with the customers pushing hardest, turning open research questions into tools their agents use, shipping the code yourself. THE FRONTIER YOU'LL WORK ON • turning an 800-endpoint enterprise API into a tool surface a model can navigate. tool granularity, composite tools, description engineering, scope design. no textbook exists; you write it in production. • building the evals that tell us an agent picked the right tool. the field has no good answer; your harnesses become ours. • working directly with the engineers defining what agents can do, and feeding what breaks back into the product. WHAT YOU'LL DO? • own enterprise accounts end to end: discovery, spec, implementation, QA, release. • ship production code in our toolkit monorepo and customer SDKs: custom and composite tools, schema and scope changes, per-customer patches. • turn large enterprise APIs into agent-usable tool surfaces, and build the evals that prove the model picks the right tool. • QA toolkits, and build the harnesses that make QA repeatable instead of artisanal. • run weekly customer syncs: the summaries, ETAs, and escalations that keep an engagement coherent. • operate per-customer tool-patching as self-service, not escalation. • dogfood Composio on your own workflows. "MUST HAVES" if you are very good, nothing is a must per-se • applied integration engineering – you ship production code, not demos. – you live in enterprise APIs and auth: OAuth 2.0, consent models, scope design, multi workspace tokens. you know the edge cases from scars. – you make architecture calls under pressure and defend them to a customer's principal engineers. • customer technical ownership – you've taken a named enterprise account from first call to live.– you turn vague asks into specs, and push back when the ask is wrong. • ai native – you have built with the language models. – you have built for the language models, and treat the agent-tool interface as an open research problem. • taste - finger feel for good developer experience. • typist - you explain complex ideas clearly under deadline. • human - you build trust and admit what you don't know. OPTIONAL • years of Python in production (Typescript a plus). • API platform, iPaaS, or integration-heavy background. • prior FDE / solutions-architect role selling to big tech. • frontier output: MCP servers, tool-use evals, agent-tooling writeups.

Member of Technical Staff – Applied AI Engineering EXPERIENCE: 2–4 years TYPE: Full-time FOCUS: Applied AI Customer-facing WORK HOURS: US Timings LOCATION: Bengaluru ABOUT THE COMPANY At Composio, we are building infrastructure that allows AI agents to communicate with the tools people use for work, including GitHub, Gmail, Notion, Salesforce, and more. We are a small team of engineers working across context, auth, execution, and reliability to build the action layer for AI agents. We raised a $25M Series A from Lightspeed, with angel investors including Guillermo Rauch (CEO of Vercel), Dharmesh Shah (CTO of HubSpot), and Gokul Rajaram. Our customers range from YC startups to companies like Brex, Glean, Zoom, and more. ABOUT THE TEAM This is an applied AI engineering role, not a traditional support role. You will work directly with customers, learn the repeated work behind their problems, and build agentic systems that take that work over. When a customer needs direct engineering, you will debug and ship. When a pattern repeats, you will turn it into a workflow, eval, tool, or product capability that compounds. • The frontier: support is one of the richest applied AI environments — an agent has to gather context, reason across systems, use tools, take bounded action, verify the outcome, and communicate clearly. You design that system. • The trust bar: AI cannot be trusted by default with customer support. You define state, permissions, guardrails, evals, and escalation paths so the automation earns trust. • The feedback loop: it's immediate. Deploy into real support work, watch where the system fails, and improve it until people can hand over more of the job. ROLES AND RESPONSIBILITIES What you'll build • Agentic workflows that take over support functions: intake, classification, context gathering, reproduction, diagnosis, response drafting, remediation, verification, and follow-up. • Tool-using agents that safely inspect runs, logs, auth state, configuration, and provider behavior, then propose or execute bounded recovery actions. • Eval suites for diagnostic correctness, resolution quality, safe escalation, customer communication, and end-to-end task completion. • Human-in-the-loop systems that make ownership, uncertainty, approvals, handoffs, and failure states explicit. • The observability, memory, and feedback loops that let these agents improve from real support cases without repeating mistakes. • Product improvements that remove entire classes of customer issues instead of handling the same symptoms faster. How you'll work • Start with the customer outcome, then work backward into the model, workflow, tooling, or product change required to deliver it. • Own the hardest issues end to end when direct engineering is needed: reproduce, isolate, mitigate, ship, communicate, and verify the resolution with the customer. • Sit with customers and partner with Support, FDE, Product, and Engineering to find work that should become software. • Move from a rough prototype to a reliable production workflow, with traces, evals, guardrails, and a clear human fallback. • Use frontier models, coding agents, and internal AI tools every day to multiply your own engineering output. • Measure success in customer outcomes improved, support work removed, and classes of failure prevented. WHAT YOU'LL BRING • Experience level: 2–4 years of professional software engineering experience. Applied AI engineering • You have built production systems with language models, tool calling, retrieval, structured outputs, multi-step workflows, or agent memory. • You know how to use models and coding agents as leverage, and where their non-determinism creates hidden risk. • You use evals, traces, failure analysis, and fast iteration to make AI systems dependable. Strong systems engineering • You are an experienced backend, platform, or integration engineer who has shipped and operated production software. • You can debug across SDKs, HTTP, OAuth, webhooks, queues, data stores, logs, and third-party APIs. • You turn an ambiguous failure into a minimal reproduction, a root cause, and a durable fix. Customer-adjacent building • You are deliberately choosing to work closer to customers because you want a tighter loop between what you build and the impact it creates. • You enjoy talking directly with technical users, turning a vague problem into a system, and watching that system change their outcome. • You communicate clearly under uncertainty, ask the exact next question, and never bluff. Workflow and product judgment • You can model a support function as states, tools, permissions, handoffs, verification steps, and escalation rules. • You know what should be automated, what should remain human, and how to increase autonomy without reducing trust. • You fix the class of problem, not only the visible symptom. OPTIONAL • TypeScript or Python in production. • Experience building AI agents, copilots, workflow automation, or eval infrastructure. • Experience with a developer platform, API product, support engineering, SRE, or incident response. • Familiarity with OAuth 2.0, webhooks, rate limits, Postgres, queues, and cloud observability. • Experience with support systems such as Plain, Zendesk, Intercom, or Salesforce. • Public technical writing, open source contributions, or unusually good debugging notes.