How it works

The agent does the work. Fixed steps, fixed checks and a person keep it right.

An AI agent is good at work that needs judgment, and unreliable at doing the same thing the same way twice. So we put it inside a written workflow, test every result with checks that never change, and leave the final word to a person.

Four steps, for any department.

  1. 1

    A person hands off work

    For example a ticket in Linear or GitHub Issues, a deviation report, or a prior authorization request.

  2. 2

    An agent does it

    It follows your written workflow, inside your own cloud account, and asks a person when something is unclear.

  3. 3

    Fixed checks test it

    Scripts that give the same answer every time. If a check fails, the agent tries again, up to a set limit, then a person takes over.

  4. 4

    A person signs

    Nothing is final without a named person's approval. Every step is recorded.

One engineering ticket, step by step.

This is how Widal.Factory runs a ticket. Each step is done by plain code, by the AI, or by a person. Only the steps that need judgment go to the AI.

  1. 1Hand-offPerson

    An engineer hands a ticket to the factory in the tracker.

  2. 2PrepareCode

    The factory makes a clean copy of the code for this ticket and gives the agent the ticket and your rules.

  3. 3Write the changeAI

    The agent reads the code, writes the change and its tests, and asks on the ticket when something is unclear.

  4. 4CheckCode

    Your tests run on the change, as on any other. If they fail, the agent sees the result and fixes the change.

  5. 5ApprovePerson

    The engineer reviews and merges. Nothing reaches the main branch without a person.

  6. 6RecordCode

    The factory closes the ticket, cleans up, and records the time and the cost in the hub.

What a check is.

A check is a small script that tests the agent's work and gives the same answer every time for the same input. It can run your tests, scan for patient data in the wrong place, or confirm that every required field of a record is filled. It has three possible answers.

Pass

The check is met. The work moves to the next step.

Not yet

The check is not met. The agent gets the result and tries again. After a set number of tries, a person takes over.

Could not check

A system the check needs did not answer. This does not count as a try, so "could not check" never turns into "passed".

A check only tests what a script can measure. Medical, scientific and business judgment stays with the person who signs. And a check is kept out of the agent's reach, so the agent cannot change the test it must pass.

Why this way, with the evidence.

Each choice answers a problem that research or real incidents have shown. Vendor studies are marked as such.

Fixed steps

The problem

An AI agent picks its own next step, so two runs of the same job can go two different ways. Anthropic's guide to building agents separates workflows, where code sets the path, from agents, where the model picks it, and recommends fixed workflows for well defined tasks. Stripe runs its internal coding agents inside fixed step lists: plain code runs the linters and pushes the change, the AI writes it, and after two failed test runs the work goes back to a person. Stripe says more than 1,300 of these changes are merged each week, each reviewed by a person.

Sources: Anthropic, Building effective agents · Stripe, Minions part 2

What we do

A written workflow sets the steps and their order before work starts. The AI fills in the steps that need judgment. Everything else is plain code that does the same thing every time.

Fixed checks

The problem

AI output looks right more often than it is right. Veracode tested over 100 models and found that 45% of the code samples failed security tests. Its 2026 report still finds an average pass rate of only 56% (studies by a security vendor). Spotify's background coding agent runs checks that the agent cannot see before it may open a pull request, and an AI judge then rejects about a quarter of the changes.

Sources: Veracode, GenAI Code Security Report · Spotify Engineering, feedback loops

What we do

Scripts test each result: your tests, security scans, or "are all the required fields filled". A failed check sends the work back, up to a set number of tries, then to a person.

Tight limits

The problem

OWASP, the open security standards group, lists "excessive agency" as the third biggest risk of AI systems in its 2026 list: an agent with more access than its task needs. In July 2025 an AI coding agent deleted a company's production database during a code freeze. OWASP's advice is least access, and a person's approval for actions with high impact.

Sources: OWASP, Top 10 for LLM Applications 2026 · AI Incident Database, Replit

What we do

Agents get no keys to the tracker or the code host. The factory acts for them. Each engineer's agents work on that engineer's own machine. Merges, and changes to production, wait for a person.

Measured

The problem

In a randomized trial by METR in 2025, experienced developers took 19% longer with AI tools, while they believed they were about 20% faster. METR's 2026 follow-up found a speedup of about 18%, but could not rule out none. The studies disagree, and feelings were wrong. Google's DORA 2025 report, a survey of nearly 5,000 people, found that AI makes an organization's strengths and weaknesses bigger.

Sources: METR, randomized trial (2025) · METR, 2026 follow-up · DORA, 2025 report

What we do

The hub records every run: time, cost, review rounds and failures. The team sees what AI really changes, and fixes the workflow where it does not help.

Your data stays yours

The problem

About half of workers use AI tools their employer has not approved (49% in a BlackFog survey, 59% in a Cybernews survey). In the Cybernews survey, about three in four of them said they had shared sensitive data with those tools. Both are self-reported surveys by security vendors.

Sources: BlackFog, shadow AI research

What we do

One approved setup in your own cloud account, under your own agreements with the model provider, such as the business associate agreement HIPAA requires.

Better every week

The problem

When each person runs agents on their own laptop, a mistake is seen by that person only, and repeats on the next laptop.

What we do

One shared set of workflows, rules and checks. Each week we review what the agents got wrong and turn it into a rule or a check, for every agent at once.

Works with the tools you already use.

Agents work in your own systems, not in a new one. A dot marks what works on day one. Our engineers connect the rest during setup.

Trackers

  • Linear(works on day one)
  • GitHub Issues(works on day one)
  • Jira
  • Azure Boards
  • ServiceNow
  • Asana

Code and reviews

  • GitHub(works on day one)
  • GitLab
  • Bitbucket
  • Azure Repos

Builds and tests

  • GitHub Actions(works on day one)
  • GitLab CI
  • CircleCI
  • Jenkins
  • Buildkite

Coding agents

  • Claude Code(works on day one)
  • OpenAI Codex
  • Gemini CLI

AI models

  • Anthropic Claude(works on day one)
  • Amazon Bedrock
  • OpenAI
  • Google Gemini
  • Azure OpenAI

Cloud

  • AWS(works on day one)
  • Google Cloud
  • Microsoft Azure

Sign-in

  • Google(works on day one)
  • Okta
  • Microsoft Entra ID
  • Any OpenID Connect provider

Chat and alerts

  • Slack
  • Microsoft Teams
  • PagerDuty

Monitoring

  • Amazon CloudWatch(works on day one)
  • Datadog
  • Sentry
  • Grafana

Documents

  • Confluence
  • Notion
  • SharePoint
  • Google Drive

Quality and regulatory

  • Veeva Vault
  • MasterControl
  • TrackWise Digital
  • Greenlight Guru

Health records

  • Epic
  • Oracle Health
  • athenahealth
  • FHIR and HL7v2 systems

The words we use.

Agent
An AI program that does a task, such as writing code or drafting a report.
Workflow
The written steps for a job, in order, that the agent must follow.
Check
A fixed test of the agent's work that gives the same answer every time.
Sign-off
The point where a named person approves, as your procedures require.
Done check
The last step of a workflow. It tests the proof that the work is finished, such as a merged change or a filed record, and keeps it with the work.
Widal.AS (AgentSystem)
AI agents, written workflows, fixed checks and sign-offs for every team, in your own cloud, with one hub for the company. Widal.Factory is its part for engineering teams.
Hub
One page for the whole company: what agents are doing, and what waits for whom.
Factory
An engineer's own machine in your cloud, where coding agents work on that engineer's tickets.
Pull request
A proposed change to the code, which a person reviews before it is merged.
Landed
A ticket whose change was merged and that was closed as done. The hub counts speed and quality on these.

New to this? Start with What is an AI software factory?

Tell us which work you want to hand off first.

Book a call