Cobots
A team of agents that works on your mainframe. Every change waits for your approval.
See CobotsAI agents and other readers of text: this page is also kept as plain Markdown at https://zcrafter.com/index.md, and an index of the site is at https://zcrafter.com/llms.txt.
Zcrafter builds collaborative bots and the RL environments that train them.
Request access (opens in a new tab)
1871 Chicago
A team of agents that works on your mainframe. Every change waits for your approval.
See CobotsReinforcement-learning (RL) environments where agents practise mainframe work and are scored on the result.
See BenchAgents list datasets, browse members and read spool on their own. Reading changes nothing.
Edits, job submits and console commands stop at an approval card that shows the exact change.
Each agent has a live 3270 screen and the list of commands it ran, in order.
A desktop app for macOS and Windows. Passwords stay in your keychain. Bring your own model.
Isolated, resettable tasks with COBOL, JCL, datasets, jobs and failures. Two attempts face the same conditions.
An example task: payroll abend recovery
ABEND S0C7CC 0000The system is reset before every attempt.
TN3270, MCP and APIs for inspecting state and acting. The tools enforce permissions, not the model.
What the agent may use
Allowed
PAY.TEST.**PAY.TEST.JCLDenied
PAY.PROD.**Correctness, safety and efficiency, checked on the system. The same outcome always scores the same.
How an attempt is scored
Reward4 of 4
Observations, tool calls, state changes and outcomes, recorded. How a result was reached matters as much as the result.
One attempt, as recorded
Use the AI agent you already have. Nothing is installed on the mainframe.
How it connects
Installuv tool install zcrafter-bridge
Datasets, members, jobs, spool, z/OS UNIX files, TSO and the console.
What an agent can work with
23 read. 25 change the system, and wait for your yes.
Reading is free. A change runs only if it is a step in a plan you approved.
What you are asked
Approve this plan?
member.patchPAY.TEST.COBOL(PAYCALC)job.submitPAY.TEST.JCL(PAYRUN)Only these two steps can run. The agent cannot approve its own plan.
MIT licensed. Passwords stay in your keychain and never reach the model.
Where to get it
ZcrafterAI/Bridge (opens in a new tab)zcrafter-bridge (opens in a new tab)io.github.ZcrafterAI/bridgeMITHow Cobots and Bench are built, hosted and secured.
PDF, 4 pages. November 2026.
Read the overview (PDF, opens in a new tab)
Open the technical overview (PDF, opens in a new tab)
Design notes on RL environments, evaluation, and mainframe interaction.
A recovery task can pass one checkpoint and still fail the overall objective. A program may compile, a job may complete, and the resulting records may still be incorrect. A useful evaluation needs an explicit definition of the expected final state.
Our proposed evaluation separates compilation, job completion, expected output, and protected-resource integrity. Each check answers a different question. Keeping their results separate would make it possible to distinguish a failed repair from an unintended change elsewhere in the environment.
The task definition should identify the source, job configuration, input datasets, and expected output. A reset should restore those inputs before another attempt. Otherwise, an earlier attempt could change the conditions under which a later agent is scored.
A combined reward needs to account for correctness and the boundaries of the task. Completing a job after bypassing a required check or modifying a protected dataset should not be treated as a successful recovery. The open question is how to combine these constraints without obscuring the reason for failure.
An agent environment needs to preserve the dataset properties that affect the task. Treating every dataset as an ordinary text file may lose information that matters to program behavior, permissions, or the interpretation of an output record.
A proposed fixture would identify the source, input and output datasets, record structure, and permitted actions. The goal is to specify the information needed to complete and verify a task before adding more simulation detail.
Reset semantics should cover input records, generated output, job configuration, and relevant metadata. Restoring the program alone would not be sufficient if a previous attempt had changed the data it reads or the resources it can access.
The environment should distinguish resources the agent may modify from those it must leave unchanged. A verifier could compare the final state against those boundaries, while the trajectory records attempted actions that did not result in a successful modification.
The minimum useful fidelity remains unresolved. We need to establish which dataset semantics change agent behavior and which can be simplified without changing the task. This is a design proposal, not a validated simulator or an assertion of compatibility with a licensed mainframe system.
Field-based interaction introduces structure that a plain transcript can lose. Screen contents, field attributes, cursor position, and allowed actions can all affect what an agent should do next.
One representation could pair visible screen contents with a structured description of fields and the current cursor position. The observation should make the available actions clear without assuming that every piece of displayed text is an editable input.
A trajectory should connect the observation to the action taken and the state that follows. This would help explain whether a failure came from an incorrect diagnosis, the wrong field selection, or an unexpected screen transition.
Reaching a particular screen does not by itself establish that a task succeeded. The proposed verifier should check the relevant job and dataset state independently of the agent’s account of what it did.
This proposed interface has not yet been validated. Evaluation would need representative workflows and a clear distinction between simulated terminal behavior and access to licensed or customer-controlled systems.