AI-generated code still needs to satisfy requirements, integrate with existing systems, and survive testing. For engineering teams, the useful question is how agents participate in that process and who owns the resulting changes.
This discussion draws on GeekyAnts’ explanation of its Agentic Development Life Cycle, alongside published approaches from Thoughtworks, EPAM, and Globant.
The companies below represent different approaches to agentic engineering. Their inclusion reflects documented capabilities, rather than an independently verified performance ranking.
What changes when agents participate across development?
GeekyAnts describes ADLC as a model where agents support planning, implementation, testing, documentation, and analysis. Engineers retain responsibility for architecture, security, quality, and release decisions.
The article also describes an ongoing engagement that connects development with post-release improvement.
The technical implication is that teams need explicit boundaries around agent work. Permission to generate a patch, execute tests, merge changes, and deploy software should be considered separately.
Conventional development already includes automation and review gates. An agentic approach therefore needs to demonstrate how it changes execution, validation, and exception handling.
Which companies offer relevant approaches?
1. GeekyAnts
GeekyAnts’ ADLC description focuses on agent participation within ongoing product engineering, supported by human checkpoints.
The article explains the intended process but does not provide comparative evidence that it outperforms other delivery models.
Useful evaluation: A feature walkthrough showing the requirement, agent-generated changes, reviewer corrections, test results, and release approval.
2. Thoughtworks
[Thoughtworks’ AI/works])combines reverse engineering, requirements enrichment, specification development, code generation, runtime operations, and governance.
Its scope makes it relevant to projects where understanding an existing system is a substantial part of implementation.
Useful evaluation: How engineers validate recovered business rules and detect behavior that an automatically generated specification misses.
3. EPAM
[EPAM’s AI/Run offering] addresses agentic workflows across the product lifecycle, alongside governance, change management, and performance measurement.
It is relevant to organizations evaluating adoption across several engineering teams.
Useful evaluation: How the approach integrates with existing repositories and delivery pipelines, and whether pilot results remain repeatable across different codebases.
4. Globant
[Globant CODA] provides an agentic software development suite covering activities including coding and testing.
Its relevance lies in coordinating AI tooling across multiple development tasks.
Useful evaluation: Whether generated tests independently check business requirements or simply reinforce assumptions already present in generated code.
A practical way to compare the approaches
A shared pilot brief would make the comparison more meaningful than separate vendor demonstrations.
Consider a hypothetical task: adding a new API-backed feature to an existing application. During implementation, the requirement changes, a dependency becomes unavailable, and a validation check fails.
The assessment should examine:
- Context handling: Does the agent identify conflicting requirements?
- Permission boundaries: Which actions require approval?
- Validation: What evidence accompanies the proposed change?
- Escalation: When does the agent stop and involve an engineer?
- Delivery effort: How much review and rework remains?
Elapsed delivery time, reviewer effort, defects, and operating costs provide a stronger basis for comparison than generated code volume.
What should “human oversight” mean?
A review checkpoint only works when the reviewer can understand and challenge the proposed change.
For a database migration, that could include the intended schema change, affected dependencies, validation results, and recovery approach. A generic approval prompt provides little basis for a decision.
The engineering standard should remain understandable changes, reproducible checks, and clear ownership after deployment.
Discussion: Which evidence would make an agent-generated change ready for review, and which decisions should remain exclusively with engineers?
Top comments (0)