DevAgent

An evidence-driven engineering agent for software engineering, offline PLC engineering/FAT review, and read-only onsite commissioning.

From requirement to evidence-backed engineering decisions.

DevAgent is free during public beta. If it saves you engineering time, consider supporting continued development.

• Project: https://github.com/tomha85/devagent
• Documentation: https://github.com/tomha85/devagent#readme
• Issues: https://github.com/tomha85/devagent/issues

General Architecture

DevAgent has three sibling product branches under one evidence-driven core. Each branch owns a distinct engineering responsibility, input model, safety boundary, and qualification path.

                         DevAgent Core
              Evidence-Driven Engineering Platform
                              |
          +-------------------+-------------------+
          |                   |                   |
          v                   v                   v
 Software Engineering     DevAgent PLC        DevAgent Live
 Product Branch #1        Product Branch #2   Product Branch #3
          |                   |                   |
 Local / GitHub repo      PLC export           Engineering context
          |               Siemens              + OPC UA endpoint(s)
          |               Rockwell                  |
          |               Schneider                 v
          |                   |              Reconcile / Trust
          v                   v              Freshness / History
 Understand / Plan       Canonical PLC               |
 Modify / Test / Review  Engineering Model           v
          |                   |              Deterministic
          v                   v              Commissioning Diagnosis
 Engineering Report      Analyze / Verify             |
 + Safe Branch           FAT / Release Readiness      v
                                                Evidence / Explanation
                                                + Next Safe Check

Read-only PLC to Live integration contract

DevAgent Live is not a child of DevAgent PLC. Stable PLC engineering context can cross into Live through a bounded read-only adapter contract. FAT and release-readiness authority stays with DevAgent PLC; OPC UA sessions, runtime trust/history, diagnosis, and commissioning Q&A stay with DevAgent Live.

Canonical PLC Engineering Model
              |
              | READ-ONLY ENGINEERING CONTEXT
              v
       DevAgent Live Adapter <----- OPC UA Runtime Evidence
              |
              v
      Engineering <-> Runtime Join
              |
              v
      Commissioning Diagnosis

The read-only engineering-context path does not transfer FAT authority, release-readiness authority, or PLC control authority to DevAgent Live.

For the expanded architecture, see: https://github.com/tomha85/devagent/blob/main/docs/general-architecture.md

Why DevAgent?

Many AI engineering tools optimize first for generation. DevAgent is built around a different question:

Can this engineering claim be supported by evidence on the exact analyzed revision, PLC export, or trusted runtime state?

Core principles include:

• evidence before modification;
• explicit acceptance contracts;
• false-VERIFIED resistance;
• minimal-change discipline;
• backups before edits;
• repository-native verification;
• independent review;
• bring-your-own-model support;
• deterministic publication rules;
• a strict read-only commissioning boundary for DevAgent Live.

Install

Standard package:

python -m pip install devagent-ai

With OPC UA runtime support for DevAgent Live:

python -m pip install "devagent-ai[live]"

Or with pipx:

pipx install devagent-ai

The PyPI distribution is devagent-ai; the Python package and CLI command are both devagent.

Software Engineering

Run DevAgent from a repository and describe the engineering outcome:

cd my-repo
devagent "Fix the websocket reconnect bug and add regression tests"

A normal software-engineering path is:

DISCOVER / UNDERSTAND
        |
        v
COMPILE ACCEPTANCE CONTRACT
        |
        v
BASELINE / PLAN / GATHER CONTEXT
        |
        v
IMPLEMENT MINIMAL PATCH
        |
        v
TARGETED + BROAD VERIFICATION
        |
        v
INDEPENDENT REVIEW
        |
        v
FINAL CURRENT-REVISION VERIFICATION
        |
        v
FULL ENGINEERING REPORT
        |
        v
VERIFIED ONLY: COMMIT + FAST-FORWARD PUSH SAFE BRANCH

DevAgent runtime does not create pull requests, merge, rebase, force-push, or deploy.

DevAgent PLC

DevAgent PLC performs offline/pre-site engineering review for supported exports from:

• Rockwell Studio 5000 / Logix Designer - full-project .L5X;
• Siemens TIA Portal - supported exported source/XML engineering artifacts;
• Schneider Electric EcoStruxure Control Expert / Unity Pro - supported XML exchange exports, with .XEF preferred.

Example:

devagent plc ./exports/Line1_Controller.L5X \
  --requirements ./requirements/line1.md \
  --ai \
  --output-dir ./plc-review/line1-current

The PLC workflow can provide engineering review, requirements verification, risk detection, FAT planning, regression analysis, recommendations, evidence, and release-readiness reporting. Static analysis does not pretend to replace simulator, HIL, field wiring, process physics, or real-controller execution when those are required.

Full PLC guide: https://github.com/tomha85/devagent/blob/main/docs/plc-engineer-guide.md

DevAgent Live

DevAgent Live is the separate read-only onsite commissioning branch. It can combine supported engineering context with trusted OPC UA runtime evidence to diagnose blockers and explain the next evidence-backed check.

Supported PLC engineering context
              +
      OPC UA runtime endpoint
              |
              v
         DevAgent Live
              |
              v
  reconcile engineering <-> runtime
              |
              v
 deterministic diagnosis + evidence

Example:

devagent live assist \
  --project-folder /path/to/customer-line1 \
  --primary-project plc/Line1.L5X \
  --endpoint opc.tcp://192.168.10.20:4840/ \
  --history-seconds 900 \
  --ai \
  --provider openai \
  --model YOUR_OPENAI_MODEL

READ ONLY safety boundary

READ ONLY
no write
no force
no reset
no bypass
no upload / download
no mode change
no start / stop control

The LLM may interpret or explain evidence. It cannot turn model confidence into PLC truth or authorize machine control.

Live documentation: https://github.com/tomha85/devagent/tree/main/docs/live

Bring Your Own Model

DevAgent supports OpenAI, Anthropic/Claude, xAI/Grok, Google Gemini, and OpenAI-compatible endpoints behind the same deterministic engineering harness.

Example:

devagent setup --provider openai --model YOUR_MODEL
export OPENAI_API_KEY=...
devagent doctor --live

Provider choice can affect reasoning quality, latency, cost, and privacy characteristics. It does not change DevAgent's deterministic acceptance, verification, safety, reporting, or publication rules.

Outcome Contract

DevAgent reports bounded engineering outcomes:

• VERIFIED - required criteria are supported by admissible evidence and final verification/review passes on the current revision;
• PARTIALLY_VERIFIED - meaningful evidence exists but complete proof is unavailable;
• BLOCKED - DevAgent cannot safely understand, implement, or verify the request.

A conservative result is preferred over a false VERIFIED.

Project Status

DevAgent is beta software. Qualification claims are tied to explicit test catalogs, pinned revisions, fixtures/external oracles where applicable, and the environments actually exercised. They are not universal-correctness or market-superiority claims.

For the current implementation status, qualification details, limitations, examples, and complete documentation, see the main repository README: https://github.com/tomha85/devagent#readme

Support

DevAgent is currently free during public beta.

If DevAgent saves you engineering time, you can support continued development through GitHub Sponsors: https://github.com/sponsors/tomha85

License

DevAgent is open source under the MIT License.

Copyright (c) 2026 Tom Ha

Original repository: https://github.com/tomha85/devagent