Applied AI Engineer

About the Company

America is under sustained cyber attack. Our adversaries infiltrate our networks, steal our IP, and degrade the digital infrastructure that modern life runs on. They’ve learned—correctly—that those attacks rarely produce consequences.

Twenty was founded to change that, by making our adversaries think twice before they attack us. Our vision is American and allied primacy in cyberspace—a future where they cannot contest us, deterrence is assured, and the free world remains secure.

Founded in 2024, Twenty Technologies ( www.twenty.io ) industrializes offensive cyber operations for the U.S. and its allies. Headquartered in Arlington, Virginia, Twenty has raised $168M from Khosla Ventures, Accel, Caffeinated Capital, Friends & Family Capital, Point72 Ventures, General Catalyst, and In-Q-Tel.

Mission | On Site | Full Time | U.S. Citizenship Required / No Active Clearance Required

Role Summary

You’ll build an agent environment at Twenty, with an interaction model similar to Claude Code and Codex. Users will give it tasks, supply context, and review its work. They’ll need to see what it’s doing and be able to intervene when it gets something wrong.

Your work will center on the agent harness : the code that assembles context, executes tool calls, tracks progress, and coordinates delegated work. You’ll decide how information carries across a session, how the agent checks its results, and how it recovers when a step fails.

You’ll write production software in Python, TypeScript, or Rust and maintain it after launch. Expect to spend time investigating failed runs, testing changes to the harness, and working with users to understand where the system needs improvement.

What You’ll Do

• Build agent harnesses. Write the execution loop that calls the model, runs tools, records results, and determines the next step. Set time and cost limits. Define stopping conditions and detect runs that repeat work without making progress.

• Manage context. Choose what the model sees at each step, including instructions, retrieved records, previous actions, and tool results. Build retrieval and summarization into the workflow. Check that summaries preserve task requirements, source references, and unfinished work. Investigate failures caused by missing information or conflicting instructions.

• Orchestrate workflows. Combine model calls with application code and human review. Implement branching, parallel execution, and checkpoints where the task requires them. Store task state so work can resume after an interruption. For delegated work, define what each agent receives, what it returns, and how shared state is updated.

• Build tool interfaces. Define inputs, outputs, and error responses. Validate arguments before execution and check results before passing them to the next step. Handle retries so a lost response doesn’t cause duplicate writes or messages.

• Build the controls people need to work with an agent. Show progress and make tool calls and results inspectable. Let users correct instructions, approve actions, stop a run, and resume unfinished work. Make clear what has completed, what failed, and what still needs review.

• Handle failures and control access. Detect invalid tool calls, dropped requirements, and reports of success that don’t match the actual result. Implement retry limits, cancellation, and escalation to a person. Enforce permissions in application code and require approval for actions that need review.

• Evaluate changes. Build test cases from representative tasks and production failures. Check results against the task’s requirements. Measure completion rate, failure severity, human intervention, latency, and cost. Run regression tests when changing models, prompts, tools, retrieval, or orchestration.

• Operate the system. Keep traces that show what the model saw, which tools ran, and how task state changed. Redact secrets and restrict access to sensitive traces. Version prompts and workflow code, test releases on a limited set of traffic, and support rollback.

Who You Are

You investigate a failed run before proposing a fix. You’re comfortable reading backend code and tracing a problem through several services. You can explain why you chose an approach, show the results of testing it, and describe what remains uncertain.

You work directly with domain experts to understand the task and its failure cases. You ask which errors are tolerable, which require review, and what a user needs to see before accepting a result.

Must Have

• Production software experience. At least four years building and maintaining production software, including hands-on work on an LLM application that calls tools or performs multi-step tasks. Be prepared to discuss failures you encountered after launch and how you addressed them.

• Programming and backend engineering skills. Strong production experience in at least one of Python, TypeScript, or Rust, with the ability to work in the others as needed. You can build services, work with asynchronous execution and external APIs, manage persistent storage, and write automated tests.

• Harness and orchestration experience. You’ve written or substantially changed an agent’s execution loop. You can explain what the model controls, what the application enforces, and when the user needs to make a decision. You’ve investigated failures involving tool use, lost context, or work that stopped before the task was complete.

• Context engineering experience. You’ve worked on retrieval, context assembly, summarization, or memory for an LLM application. You can explain how you chose the information supplied to the model, worked within token limits, and tested whether the right information survived across steps.

• Experience handling partial failure. You understand timeouts, retries, concurrency, and idempotency. You can design recovery for a workflow that fails halfway through and verify whether an action completed before attempting it again.

• Evaluation and debugging experience. You’ve built tests for model-driven workflows and used execution traces to investigate failures. You can choose checks suited to the task, account for variation between runs, and compare automated grading with human assessments.

• Security knowledge. You understand prompt injection and the risks of instructions embedded in retrieved documents or tool output. You can restrict tool access, protect credentials, and implement validation and approval steps for actions such as deleting records, exporting data, or sending messages.

Nice to Have

• Experience with long-running or multi-agent workflows, including checkpoint recovery, task delegation, and shared state.

• Experience building retrieval or memory systems with source tracking, freshness checks, and access controls that follow the requesting user’s permissions.

• Experience running controlled comparisons or ablation studies to measure how individual components affect results.

• Experience with self-hosted inference, inference optimization, or task-specific fine-tuning, including measuring their effect on task performance and operating cost.

Tech Environment (You Might Work With)

Python, TypeScript, Rust, model APIs and SDKs, relational databases, object storage, search systems, asynchronous workers, workflow engines, containers, and tracing and evaluation tools.

The implementation may include an orchestration framework or a custom harness. Be prepared to read and debug the execution code, including framework code the application depends on.

Benefits

What's on the table:

• Health. Medical, dental, and vision plan options. Life / AD&D, disability coverage options.

• Family. Paid parental leave for eligible full-time employees. 12 weeks for birthing parents, 4 for non-birthing parents, 6 weeks for adoptive, foster, or intended parents through surrogacy.

• Vacation. Paid holidays and flexible PTO. Take what you need.

…