OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview

Executes shell commands in a sandboxed terminal environment while maintaining bidirectional context with an LLM agent. The agent receives command output, error streams, and exit codes in real-time, enabling it to reason about execution results and decide on next steps. Implements a command-response loop where the LLM can chain multiple commands based on previous outputs, with built-in handling for interactive prompts and long-running processes.

Breaks down complex terminal-based tasks into executable subtasks using chain-of-thought reasoning. The agent generates a plan, executes steps sequentially, and dynamically adjusts the plan based on intermediate results. Implements backtracking logic where failed steps trigger re-planning with updated context about what went wrong.

Enforces a schema-based constraint system where the LLM can only execute actions (commands, API calls) that conform to predefined schemas. The framework validates action parameters before execution, preventing malformed or dangerous commands from reaching the terminal. Implements a registry pattern where actions are registered with type hints, constraints, and execution handlers.

Maintains a structured history of all executed commands, their outputs, and side effects. The agent can query this history to understand what has already been done, avoiding redundant operations. Implements state snapshots at key points, allowing the agent to reason about system state changes and detect when commands had unexpected effects.

Automatically detects command failures (non-zero exit codes, timeout, resource exhaustion) and implements retry strategies with exponential backoff. Different error types trigger different recovery strategies: transient errors retry immediately, resource errors wait before retrying, and permanent errors trigger re-planning. Includes timeout handling for long-running commands with configurable thresholds.

Abstracts the LLM backend behind a unified interface, allowing the agent to work with different providers (Gemini, OpenAI, Anthropic, local models) without code changes. Implements provider-specific adapters that handle differences in API formats, token counting, and function-calling schemas. Supports model switching at runtime based on task requirements or cost optimization.

Implements instrumentation and metrics collection throughout the agent execution pipeline to identify bottlenecks. Tracks latency per component (LLM inference, command execution, planning), token usage, and task success rates. Provides hooks for performance profiling and optimization, with built-in support for A/B testing different strategies.