DevOps Engineering Agent¶

Source: agent_library/agents/devops_gym_agent

What This Agent Does¶

devops_gym_agent is a general-purpose software-engineering agent aimed at the kind of work a junior DevOps engineer does every day:

• Build & configuration: fix build failures, resolve dependency issues.
• Monitoring: diagnose runtime anomalies using standard system tools (top, htop, ps, watch, curl, df, netstat, lsof).
• Issue resolving: generate patches that fix GitHub-style issues.
• Test generation: write tests that reproduce the bug and validate the fix.

It sits one step up from harbor_agent: the tool palette is the same terminal + file-ops set, but the agent now has access to dynamic sub-agent orchestration primitives (create_subagent / call_subagent / list_subagents) and a think tool for explicit planning. No debugger sidecars, no graph-database memory; the premise of DevOps-Gym is that the raw Linux terminal plus a planning loop is enough.

Key Design¶

• think as a first-class tool. The workflow in the system prompt explicitly tells the agent to plan before acting.
• Fresh-perspective sub-agent for deadlock. When stuck after several attempts, the agent is instructed to spin up a fresh sub-agent (create_subagent + call_subagent), optionally on a different model, for an unbiased second pass.
• Pattern-specific completion. For monitoring tasks, the agent writes a canonical /workspace/monitor_result.txt with the anomaly pattern and the reason: a contract the evaluator reads.
• No evaluation scripts. The prompt explicitly forbids reading or executing grading scripts in the environment.

Agent Source¶

from opensage.agents.opensage_agent import OpenSageAgent
from opensage.toolbox.finish_task.finish_task import finish_task
from opensage.toolbox.general.agent_tools import (
    complain, think,
)
from opensage.toolbox.general.bash_tools_interface import (
    get_background_task_output,
    list_background_tasks, run_terminal_command,
)
from opensage.toolbox.general.fileop import str_replace_edit, view_file
from opensage.toolbox.general.orchestration_tools import (
    call_subagent, create_subagent, get_available_models, list_subagents,
)

def mk_agent(opensage_session_id, model=None):
    if model is None:
        model = LiteLlm(model="openai/gpt-5.5")
    return OpenSageAgent(
        name="devops_gym_agent",
        model=model,
        instruction=_SYSTEM_PROMPT,
        tools=[
            finish_task,
            think, complain,
            view_file, str_replace_edit,
            run_terminal_command,
            list_background_tasks, get_background_task_output,
            get_available_models,
            create_subagent, list_subagents, call_subagent,
        ],
    )

See the source for the full _SYSTEM_PROMPT with per-task guidance (build, monitoring, issue-resolving, test-generation).

Run It¶

uv run opensage web \
  --agent agent_library/agents/devops_gym_agent \
  --config agent_library/agents/devops_gym_agent/config.toml \
  --port 8000

The example ships a Dockerfile in the agent directory used to build the target image for the sandbox.