Case 01 / Agentic engineering

t0rque

It started with an engineering-capacity problem, not an AI demo.

What if an engineering system could take a real piece of work from ticket context to an inspectable pull request—without pretending human judgment had become obsolete?

Inspect the designed system
01problem

The small work was consuming the important work.

Small upgrades, client-specific configuration changes and bug fixes were consuming engineering capacity that should have gone into architecture, product improvements and new features. Even with client behaviour designed as modular database configuration, the volume of changes stretched the team far enough that feature delivery sometimes moved into weekends.

The code change was often not the expensive part. A developer first had to reconstruct the application: dependencies, blockers, coding standards, files and methods distributed across the codebase. Missing one relationship could turn a locally correct change into a regression elsewhere.

02changed

Then the capability boundary moved.

Tool-using agents made a longer workflow plausible: inspect a ticket, recover project context, map the relevant repository surface, propose a plan, wait for approval, perform the work and return evidence.

The opportunity was broader than Turb0. The product idea was a virtual employee that could eventually take on bounded roles across the delivery cycle. The implemented proof focused on software-engineering work, where the complete loop could be tested without pretending every organizational role already existed.

03decision

Build the workflow, not a code-generating party trick.

t0rque was conceived as an engineering system that meets teams inside Jira, repositories and pull requests. The unit of value is not a clever response. It is responsibly moving a piece of work forward.

That decision shaped the architecture: a Python body for deterministic actions, a Hermes Agent brain for reasoning and tool use, persistent project memory, an inspectable plan, explicit human approval and evidence after execution.

Activity is easy to automate. Accountability is the design problem.
04built

A working implementation made the argument concrete.

Akshay conceived the product, designed the system and built an active working prototype. It reads Jira work, retrieves and updates product context, inspects repositories, produces a plan, waits for approval, creates a branch, changes code, runs available local validation, reviews the result, captures browser evidence, opens a pull request and updates Jira.

A deliberately small public-safe demonstration changed a dashboard heading in a todo application. The task was ordinary by design; the test was whether the system would preserve the full engineering workflow around it rather than jumping from prompt to code.

05judgment

Human approval is part of the architecture.

Before repository work begins, a person inspects the proposed plan and explicitly authorizes development. A second human boundary remains after proof of work, when the pull request is reviewed.

This is not an apology for incomplete automation. It is a deliberate boundary around consequential changes, uncertain context and the difference between producing code and accepting responsibility for it.

06after

The demonstration earned a broader question.

Invoke leadership was impressed by the working demonstration and wanted to explore t0rque across teams. That expansion is currently on hold while production-readiness work remains unfinished and Akshay is committed to other delivery priorities.

The prototype still establishes the operating model: autonomy where repetitive work benefits from it, explicit control where judgment changes the consequence, and traceability across the whole piece of work.

Ambition is useful. An accurate status line is more useful.

The virtual employee, without the metaphor doing all the work

A body that acts. A brain that reasons. Memory that compounds.

t0rque separates external actions from agentic judgment, then keeps a person at the two boundaries where permission and acceptance matter.

01 / Body

Python handles the outside world.

The deterministic body performs the actions: reading and writing Jira, preparing repositories, creating branches, running commands, capturing browser evidence, pushing work and opening pull requests.

Hands and legs, not judgment theatre.
02 / Brain

Hermes reasons across the work.

The open-source Hermes Agent provides the reasoning and tool-using layer used to interpret context, form a plan, generate and review code, and decide which action should happen next.

Reasoning remains bounded by tools, permissions and human gates.Inspect Hermes Agent
03 / Memory

The project context outlives one task.

Completed Jira work, product knowledge and human feedback update persistent project memory so later tasks begin with more context. This is memory and context maintenance—not model retraining.

Previous work becomes usable context instead of disappearing into ticket history.
04 / Judgment

People authorize consequence.

A person approves the implementation plan before repository work begins and reviews the proof of work and pull request afterward.

The system can prepare a decision. It cannot quietly grant itself permission.

A controlled end-to-end proof

The change was intentionally small. The workflow was not.

NAP-8 asked for one dashboard heading to change. That made it possible to inspect whether t0rque preserved context, scope, approval, isolation, verification and communication around an ordinary piece of work.

  1. 01

    Jira request

    NAP-8 / Change the header

    • Change the app header from ‘Todo Dashboard’ to ‘Task Manager’.
    • Repository: todo-app
    • Acceptance: the dashboard heading displays ‘Task Manager’.

    A small, controlled task made the workflow easy to inspect without exposing proprietary code.

  2. 02

    Proposed plan

    Scope before files change

    • Locate the dashboard heading and related tests or snapshots.
    • Change only the user-facing heading.
    • Avoid unrelated styling, layout, behaviour, dependency or configuration changes.
    • Inspect available build and test commands; surface unknowns before execution.

    The plan named risks and unknowns, then stopped for the Jira command ‘@T0rque approve plan’.

  3. 03

    Isolated work

    One branch. One intended change.

    • Repository: /workspace/todo-app
    • Base branch: dev
    • Working branch: t0rque/fix/NAP-8-change-the-header
    • Commit: 649eaaa

    The resulting diff changed the Dashboard heading from ‘Todo Dashboard’ to ‘Task Manager’.

  4. 04

    Proof and review

    The work returned to the team.

    • Browser verification captured the updated heading.
    • Pull request #2 targeted the configured dev branch.
    • Jira moved from In Progress to Needs Review with the branch, PR and proof-of-work context.
    • A person retained pull-request review authority.

    The PR was deliberately closed so the demonstration could be run again. GitHub-hosted CI checks were outside the prototype.

Public-safe transcript reconstructed from the supplied Jira and GitHub demonstration captures. No production customer code or client data is shown.

The designed system

One job. Ten inspectable hand-offs.

Select a stage to see what it must determine, what it leaves behind and where a plausible answer can still be wrong.

09 / 10

Human decision boundary

GitHub pull request

Return the work to the team’s normal review process.

Goes in

  • Reviewed changes
  • Evidence
  • Implementation summary

Trying to determine

  • Is the work ready to enter team review?
  • What context must accompany it?

System responsibility

  • Create a pull request with an inspectable explanation.
  • Preserve the human review and merge boundary.

Comes out

  • Pull request
  • Review context
Where confidence can become dangerous

A pull request is a review request, not proof of correctness.

Human control

The team retains review, change-request and merge authority.

Read the complete system as text
  1. Jira intakeBegin with a real unit of requested engineering work.
  2. Context retrievalGive the system enough product and code context to reason usefully.
  3. Repository mappingFind the code boundaries and dependencies the change may cross.
  4. Planning agentTurn context into an inspectable implementation proposal.
  5. Human approvalStop consequential work until a person accepts the proposed direction.Human decision boundary
  6. Branch and implementationExecute the approved plan inside the relevant codebase boundaries.
  7. Self-reviewInspect the work before asking another human to trust it.
  8. Evidence captureMake the result inspectable rather than merely reporting completion.
  9. GitHub pull requestReturn the work to the team’s normal review process.Human decision boundary
  10. Jira updateClose the communication loop where the work began.

What blocks broader rollout

A working loop is not yet a production platform.

The next phase is mostly operational engineering: isolate environments, schedule concurrent work economically and keep model usage bounded.

  1. 01Environment boundaries

    Projects that require several client-specific VPN contexts—particularly native Windows automation—cannot be handled cleanly inside one current agent environment.

  2. 02Concurrency

    The current architecture would need separate EC2 execution capacity for concurrent jobs rather than efficiently scheduling several isolated runs inside a production platform.

  3. 03Cost control

    Operating cost varies with model usage and task length. Production use needs explicit model-routing, budgets and usage controls rather than an open-ended LLM bill.

What the project actually demonstrates

Not a list of tools. A sequence of judgments.

  1. 01Notice a recurring operational constraint.
  2. 02Recognize when new technology changes the feasible answer.
  3. 03Turn the opportunity into a product and system design.
  4. 04Build enough of the difficult path to test the argument.
  5. 05Keep human judgment visible where consequences matter.
  6. 06Describe the outcome without promoting interest into adoption.