How 5 Companies Use AI Coding Agents and Assistants Across the SDLC
An agent-generated pull request has a green build and a full set of tests. A reviewer compares it with the ticket and discovers that the agent guessed how duplicate requests should behave. Every generated test uses the s
An agent-generated pull request has a green build and a full set of tests. A reviewer compares it with the ticket and discovers that the agent guessed how duplicate requests should behave. Every generated test uses the same guess.
The team can still fix the change during review. An earlier opportunity was to catch the missing rule when the agent proposed its implementation plan.
Start with an invoice request that can be retried
Consider a hypothetical invoice API. The ticket says the endpoint should handle duplicate submissions. That leaves several decisions open. Should a retry return the existing invoice? What happens if the caller reuses the same idempotency key with a different amount? What happens when two requests arrive together?
The approved plan could spell out three outcomes. A request with the same key and payload returns the existing invoice ID. A request with the same key and a different payload gets a conflict response. Two simultaneous requests with the same key and payload create one invoice record.
Those decisions give the agent something concrete to implement and the reviewer a way to judge the result.
| Stage | What the agent can prepare | What the engineer can verify |
|---|---|---|
| Requirements | A draft specification and a list of open questions | The duplicate-request rules and acceptance criteria |
| Implementation | An API change, a database migration, and a branch | The database constraint, transaction boundaries, and existing behaviour |
| Testing | API, integration, and concurrency tests | Whether tests cover changed payloads and simultaneous requests |
| Merge | A pull request with a summary of check results | Which checks passed, how the change can be deployed, and who will own it |
Two sequential requests can pass even when simultaneous requests create two records. A concurrency test gives the reviewer evidence about that third acceptance criterion. The reviewer can also check whether the test exercises the same database constraint the application uses in production.
Treat a failed test as a decision point
Suppose a generated integration test fails because it expects a conflict when the same key is reused with a different amount. The implementation currently returns the original invoice regardless of the payload.
The failure could indicate a bad implementation, a faulty test fixture, or a requirement that changed after the test was written. An engineer can trace it back to the approved rule and decide what needs to change. Here, the policy already says that a different payload should produce a conflict, so the implementation needs correction.
That review matters when the same agent wrote the code and the tests. A reviewer can add an independent scenario from the approved ticket and confirm that it fails against the incorrect implementation before the fix goes in.
Five company approaches developers can compare
GeekyAnts describes agents drafting requirements and plans, implementing changes, and running tests. Engineers approve the plan before coding, decide how to handle failures, and review and merge changes from a separate branch. The engineer who merges the change owns it in production. Requirements-to-release video
Thoughtworks documents how AI/works turns requirements and existing system context into a specification, then generates implementation stories, code, and tests from it. The specification gives reviewers a place to check that a change preserves existing invoice rules.
Globant describes a Test AI Agent that drafts and executes test plans. Teams evaluating AI-assisted QA can ask how generated plans cover concurrency, unexpected input, and behaviour across service boundaries.
EPAM describes using CodeMie in its own delivery process to keep requirements and architecture connected to implementation, QA, and release. The workflow illustrates how an approved rule can stay visible as work moves from a ticket to a pull request.
Accenture collaborated with GitHub on a randomized Copilot trial with enterprise developers and also studied adoption. The research gives teams a reference for evaluating how a coding assistant fits into an existing engineering workflow.
Put the evidence in the pull request
A pull request template can make this workflow easier to repeat. Three fields are enough for the invoice example. Link to the approved acceptance criterion, show the evidence for the riskiest behaviour, and name the person who owns the change after merge.
For the concurrency rule, that could mean linking to the ticket and attaching an integration test result showing that simultaneous retries produced one invoice. A reviewer can then start with the behaviour that matters and inspect the code that protects it.
Where does an AI coding agent need to pause in your teamβs workflow?
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.