stoiclion/agent-workbench

Agent Workbench

An OpenEnv environment of short, tool-using work tasks across eight domains: finance, industrial, math, media, office, science, security, software. Image: ghcr.io/ohdearquant/agent-workbench:e59670df586e3452cf1edb757218cbb9a81d456e.

What it teaches

Habits that carry over to unfamiliar agent tasks:

• One exact JSON action per turn. Every action is {"tool": <name>, "args": {...}}. A refused call (unknown tool, unexpected argument, a path outside the workspace, a blank answer) comes back with the reason and does not end the episode; three refusals in a row do.
• Read before answering. Each task's inputs are files in a private workspace (CSV tables, logs, notes, a subtitle file, a small Python module). The instruction says which ones matter; some files hold distractors.
• Compute instead of guessing. run_python runs a script in the workspace, so totals, rates, hashes and counts can be computed exactly.
• Verify before finishing. check says whether a candidate answer is right and nothing more (twice per episode at most); code tasks have run_tests.
• Exact answer formats. Each instruction states the format (a number with N decimals, a name, a sorted comma-separated list, a timestamp as written in the file).

Tasks

tasks.jsonl has 40 tasks, five per domain, all train. Each line gives the task id, domain, template, instruction, answer kind and file names; data/<task_id>/ holds the files exactly as the agent sees them. A task is generated from its id, so every reset of a task is identical.

Domain	Templates
finance	loan, npv, outstanding
industrial	line, oee, sensor
math	divisible, paths, powersum
media	edit, headline, subtitles
office	expenses, meetings, reports
science	gas, halflife, molar
security	bruteforce, firewall, ioc
software	merge_intervals, parse_duration, rolling_mean, slugify, top_words

Tools

list_files, read_file {path, start, end}, run_python {code}, write_file {path, content}, run_tests (code tasks), check {answer} and submit {answer}. An episode allows 20 tool calls.

Reward

Only the final observation carries a reward, in [0, 1]:

reward = 0.7 x correctness
       + 0.1 if the agent read a file or ran code before submitting
       + 0.1 if it verified what it submitted: check confirmed that exact answer, or, for
         code tasks, the visible tests passed after its last change to the files and no
         file in the workspace changed between that test run and the submission
       + 0.1 if it made no refused calls

Listing the files does not count as reading them. Correctness is computed from the task's own data when the episode runs; no answer is stored in the image or here.

• Numbers: 1.0 when the value is within the stated rounding and the answer is the number alone; 0.5 when the value is right but wrapped in other text; 0 otherwise.
• Names, addresses and timestamps: 1.0 for the value alone (case-insensitive); 0 for anything else, since crediting text that merely contains the value would reward listing every candidate.
• Lists: 1.0 for the exact sorted list; 0.8 for the right items in another order; half the share found when some are missing and none is wrong; 0 for a list with any item that does not belong.
• Code: every visible example must pass and every further case the original code passed must still pass; the credit is then the share of the further cases the original code fails that the fixed function passes, each run in its own process. The further cases are not the visible examples, so code that special-cases the visible inputs earns nothing; the code submitted unchanged, or with an edit that fixes nothing, earns 0. Each case starts with none of the agent's processes running and none of its files in the shared tmp directories, and a case that changes any file in the workspace (its bytes, mode or modification time) fails, as does every case after it, so code cannot carry a count or an answer from one call to the next.
• Headline (media): 1.0 when every rule in the brief holds; 0 when the headline is over the length limit; otherwise partial credit only once the required words are present.

An episode that ends without submit scores 0: the tool budget ran out, three calls in a row were refused, or the reply was not a valid action, in which case the trainer closes the episode with the finish action. A blank answer is refused rather than scored.

Sandbox

Each episode runs agent code under its own unprivileged uid with the system interpreter, in a workspace the episode creates and removes. The workspace sits inside a directory the uid can pass through but not write, so agent code can empty its workspace but cannot rename it or put anything in its place. Concurrent episodes cannot read each other's workspaces, a call's cleanup kills only its own episode's processes, and the server's code is not readable from agent code. When an episode ends, and before each graded case, its processes are killed and the files its uid left anywhere under /tmp, /var/tmp and /dev/shm are removed; files another uid wrote are left alone, and directories the uid could not write in are not entered. When the episode ends its workspace is removed too. File tool paths are relative, stay in the workspace and may not pass through a symlink or a reserved name; the file tools read and write regular files only. The server walks agent-made trees through directory descriptors with a stack of its own, so a path longer than PATH_MAX or a tree deeper than the interpreter's recursion limit neither escapes a call as an error nor outlives the episode. Text the tools send back is valid UTF-8 whatever bytes the agent wrote; text the agent sends in must be text UTF-8 can encode.

Source

environment/ holds the server source and the Dockerfile the image is built from (OpenEnv pinned at 86a180ede21e044f7929b9a7783ad83aa67d83a3), copied from apps/agent_workbench in https://github.com/ohdearquant/lionagi. The files generated from the OpenEnv environment template keep the template's BSD-3-Clause license; the rest is Apache-2.0.