New Software factory: work item in, reviewed pull request out AI ops · across the full delivery cycle

Review the PR. Triage the telemetry. Before anyone gets paged.

Marka Agents plug AI into every stage of how software gets shipped — plan, code, review, release, observe. Built Azure-native, powered by frontier models through Azure OpenAI, and ready to run autonomous or as your second pair of eyes.

  • Currently watching
  • 0 repos
  • ·
  • 0 App Insights workspaces
  • ·
  • 0 work items this month
Built for Azure DevOps App Insights Entra ID · Runs on Azure OpenAI
Why Marka Agents

AI through every stage of delivery,
on the stack you already run.

Marka Agents is an operations layer that plugs into Azure DevOps, App Insights, and your Entra tenant, runs on frontier models through Azure OpenAI, and turns AI into measurable lift on quality and throughput, without forcing a single new vendor on your team.


02

Azure-native by design

Built first for teams already on the Microsoft stack. We speak Azure DevOps pull requests, work items, and pipelines natively — and read App Insights and Log Analytics as first-class signals, not as a generic webhook.

  • Azure DevOps
  • Application Insights
  • Log Analytics
  • Azure AD / Entra ID
  • Azure OpenAI
  • Key Vault
  • Service Bus
  • Functions
03

Runs in your Azure subscription

Marka Agents deploys from Azure Marketplace straight into your own subscription — the AI runs on an Azure OpenAI account you own, and your code, telemetry, and findings never leave your environment. Nothing is sent to us, ever.

  • Your subscription Everything deploys in your Azure Private
  • Your Azure OpenAI AI runs on an account you own Managed
  • No training Inputs never train models Assured
04

Autonomous, or your second pair of eyes

Run agents fully autonomous: they review, decide, publish, and close the loop. Or run them as an assistant — they draft, you approve. Switch per repo, per team, per severity. Either mode, the routine work disappears.

Autonomous

Reviews, decides, publishes. Audit trail per action.

Assistant

Drafts the finding. A human approves before it lands.

Software factory New · Early access

Assign a work item.
Get back a reviewed pull request.

Three agents — a planner, a worker, and a reviewer — take an item from your Azure DevOps board to a pull request with a review verdict attached. No person has to step in between. Your team still decides what merges.

  1. You

    Assign the item

    Assign a Product Backlog Item or Bug to the Marka agent on your Azure DevOps board. That is the only trigger.

    Assigned To · Marka Agent
  2. Planner

    Split it into tasks

    Reads the item and the repository, then creates up to eight child Tasks, each with a goal, files, done criteria, and checks. Thin items come back with questions instead.

    Tasks · plan comment
  3. Worker

    Write the code

    Implements the Tasks in dependency order inside an isolated sandbox, builds, runs the unit tests, and commits each Task to one branch.

    1 branch · 1 draft PR
  4. Reviewer

    Review the result

    Reviews the final commit once and posts a verdict, with findings ranked by severity, on the pull request.

    Verdict · findings
  5. You

    Approve or reject

    Your team reads one pull request with its review attached. Nothing merges on its own.

    Human merge

Why teams use it

Backlog to pull request, no hand-offs

The board is the workflow. Progress shows up as tasks, comments, and a pull request on the item your team already watches. No new tool to learn.

The model never holds your keys

Agents propose; trusted code applies. The AI has no Azure DevOps credential. Validated code publishes every commit, comment, and pull request through your safety settings.

Checked before you look

The worker builds and runs your tests before it commits. The reviewer reads the final commit before a person does. Your engineers start from a reviewed change, not a blank page.

Adds, never rewrites

The planner only adds tasks and comments. The worker never merges, never closes tasks, and never touches pipelines or infrastructure files. Every change is easy to trace and undo.

Recovers instead of repeating

Each step is saved before it runs. After a crash the factory resumes where it stopped, without a second model attempt or duplicate tasks, comments, or pull requests.

Every run in one table

The Agent Runs page lists every planner, worker, and reviewer run with its stage, pull request, and duration. Stops and completions also land as comments on the work item.

Where it fits

Best for work a developer could finish in a day. Architecture, infrastructure, and pipeline changes stay with your team — the worker is not allowed to touch them.

Request early access
  • Well-scoped backlog items

    Small features with clear acceptance criteria that wait in the backlog because nobody has a free afternoon.

  • Reproducible bugs

    A Bug with steps and an expected result becomes a fix, a test, and a reviewed pull request.

  • Routine engineering work

    Added validation, missing tests, consistent logging, small refactors inside one module.

  • Planning only

    Keep the worker off and use the planner alone: every item gets split into tasks with done criteria, or comes back with the questions that block it.

  • Overnight throughput

    Assign items before you leave. Review the pull requests when you are back.

Every agent run.
One clear view.

See which agent ran, how it started, what it produced, and where it needs attention. Follow the work item or pull request back to Azure DevOps.

Marka Agents / Agent Runs Illustrative data

Agent runs

Illustrative agent runs with fictional items and sample durations, not measured benchmarks.
Agent Trigger State Item Duration
Planner Sample duration: 32s Work item hook Succeeded
Work item #842 Prevent duplicate orders on retries
32s
Worker Sample duration: 6m 48s Chain Succeeded
Pull request #317 Prevent duplicate orders on retries
6m 48s
Reviewer Sample duration: 24s Chain Succeeded
Pull request #317 Prevent duplicate orders on retries
24s
Work item advisor Sample duration: 11s Manual UI Succeeded
Work item #856 Reject invalid CSV imports
11s
Planner Sample duration: 19s Work item hook Needs attention
Work item #861 Retry failed report exports
19s
Illustrative scenarios · fictional IDs Sample durations · not benchmarks
Planner, worker, reviewer, and advisor in one run history. Your team decides what merges.
The agents

Five agents. One console.
Autonomous, or your second pair of eyes.

Each agent runs on its own surface — plans, code, reviews, telemetry, tickets — and reports into the same console. Flip a single switch per repo or team: let agents publish on their own, or have them draft and wait for a human signature.

  1. 01

    Planner

    Turns work items into tasks

    Picks up a PBI or Bug assigned to the agent, reads the repository, and splits the item into child Tasks with done criteria. Asks questions instead when the item is too thin.

    • Azure DevOps Boards
  2. 02

    Worker

    Implements tasks in a sandbox

    Codes each Task in an isolated sandbox, builds and runs the tests, and returns commits. Trusted code checks them and opens one draft pull request per item.

    • Azure Repos
    • Container Apps
  3. 03

    PR Reviewer

    Reviews pull requests automatically

    Reads the diff, walks the call graph, and writes severity-ranked findings with suggested fixes. Posts back to Azure DevOps when you say so.

    • Azure DevOps
    • GitHub
    • Bitbucket
  4. 04

    Telemetry Analyst

    Triages App Insights signals

    Sweeps log analytics on a schedule or on demand. Distinguishes infrastructure noise from real incidents. Drafts a clean summary you can sign off in seconds.

    • Application Insights
    • Log Analytics
  5. 05

    Work Item Reviewer

    Improves bug reports before they ship

    Checks every new bug for repro steps, environment, expected vs. actual, and link integrity. Flags the gaps to the reporter, not to the engineer who'll be paged later.

    • Azure DevOps
    • Jira
  6. 06

    Operator Console

    Keeps a human in the loop

    One place to see every run, every finding, and every published item. Publish policies, dry-runs, and per-team automation limits are right where the work is.

    • Web
    • SSO
Microsoft Teams New · In use at Marka

Ask in Teams.
Skip the timesheet.

Marka Agent is also a personal chat in Microsoft Teams. At Marka, the team uses it to book time and pull reports from our staff management system in one message, without opening another screen.

Microsoft Teams / Marka Agent Personal chat
You · 15:38

book 8 hours for today with the description: Implemented a bot for Microsoft Teams

Marka Agent · 15:38

Saved for you: 2026-11-06, 8h 0m, Staff Management Platform.

Description: Implemented a bot for Microsoft Teams

Entry ID: 5430669b… Receipt uses the values sent, not a read-back.

Recreated from a real exchange at Marka Entry ID shortened
  • Book time

    One complete message saves the entry to your default project, or the one you name, and returns a receipt.

    book 8 hours for today with the description: …
  • Check your hours

    Reads your own entries for any date range, with totals per day and project.

    how many hours did I book this week?
  • See team hours

    Leads see entries on the projects they lead. Management sees every entry. Everyone else sees only their own.

    show Anna's hours for last week
  • Run monthly reports

    Filters and groups a month by person, project, and customer. Totals cover every matching entry, not a first page.

    hours by person grouped by project for the previous month
  • Review absences

    Counts vacation, sick leave, and unpaid requests by person, status, and type for a year.

    which vacation requests are pending this year?
  • Ask anything else

    Ordinary questions go straight to the model, with no sign-in and no staff data.

    summarize this error message for a client
Each person signs in as themselves. The staff system's own permissions decide what they can read. One message saves at most one entry. Trusted code validates what the model asks for. No chat history is kept. Entry comments and absence reasons never reach the model.

Running in Marka's own Microsoft 365 tenant today. Each capability has its own switch, and all of them are off by default.

PR #142 LoyaltyPlatform · feature/redemption-v2 Published · 4 of 5
5 findings / 4 eligible / 1 below threshold
Error Concurrency Lines 88–142
src/Loyalty.Redemption/RedemptionService.cs

RedeemAsync debits the customer's points before persisting the redemption record. If the insert fails — duplicate id, transient EF exception — the points are gone and the redemption never existed, with no compensating action.

Suggested fix

Wrap the points-debit and the redemption-insert in a single database transaction, or move both into an idempotent stored procedure keyed by the redemption id.

Warning API contract Lines 14–22
src/Loyalty.Api/Endpoints/RedeemEndpoint.cs

POST /redeem returns 400 Bad Request for both malformed payloads and business-rule failures. Clients can't tell the two apart, and the SDK retries 400s as if they were transient.

Suggested fix

Return 422 Unprocessable for business-rule failures (insufficient points, expired campaign) and reserve 400 for malformed payloads. Update the OpenAPI schema accordingly.

Info Naming Lines 31–34
src/Loyalty.Domain/Models/Voucher.cs

Voucher.Validity mixes two ideas — issued-at and expires-at. Splitting them clarifies intent and matches the schema in the redemption-history report.

Publish policy: Min severity = Warning · Read-only off · Dry-run off
See a full run
Operator console

One console.
Every run, every finding, every published item.

The operator console is a centralized interface where humans-in-the-loop schedule analyses, inspect findings, set publish policies, and watch system health. SSO, audit trail, configurable automation limits — and every AI-assisted action logged with full provenance.

Control & safety

Automation with the lever
always within reach.

Three switches decide what reaches an engineer's inbox. None of them are global toggles — they are per-team, per-repo, per-workspace.

Read-only

Active

Marka Agents read, rank, and present — without posting a comment, editing a work item, or calling a webhook. Perfect for the first two weeks.

publish.readOnly: true

Dry-run

Per-team

Runs end-to-end including the publish step, but every outbound call is rendered to a log instead of sent. Reviewable, replayable, never visible to authors.

publish.dryRun: truepublish.dryRun.log: azure-blob://...

Minimum severity

Warning

Set the floor for what publishes automatically. Errors post on their own, warnings post on their own, info findings stay in the console.

publish.minSeverity: Warning

Built for enterprise adoption from day one.

SSO
OIDC + SAML
Tenancy
Single-tenant
Data
Stays in your subscription
Audit
Every publish, signed
Models
Configurable per tenant
Network
Private endpoints
Get in touch

Bring Marka Agents into one team.
See what they catch in a week.

A 20-minute walkthrough, then a sandboxed deployment in your tenant. Read-only by default — nothing is published to your engineers until you say so.

Or write to info@marka-development.com — we usually reply within a working day.