Agentic AI is starting to change a basic assumption in software development: AI no longer has to stop after generating an answer.
A generative AI tool can suggest code, summarize research, or create a design concept. An AI agent can potentially take that output, use tools, perform actions, evaluate results, and continue working toward a goal.
That distinction has important implications for product teams.
A recent panel discussion at thegeekconf Mini, featuring technology and product leaders from GeekyAnts, Shell, and KPMG, explored what this transition could mean for engineering, product management, design, QA, and the traditional software development lifecycle.
The most interesting takeaway was not that AI will eliminate product teams. It was that the unit of work inside those teams is changing.
The full discussion is available here:
Looking at that discussion alongside developments across the software engineering industry, here are five companies worth watching for their work around agentic AI and AI-powered product engineering.
What Changes When AI Moves From Generating to Acting?
One of the simplest distinctions raised during the panel was that generative AI primarily creates something, while agentic AI can potentially close a workflow.
That could mean an agent:
- researches a problem
- chooses an appropriate tool
- generates an output
- executes an action
- checks what happened
- corrects an error
- continues toward the intended result
For development teams, this is significantly different from autocomplete.
It starts affecting the entire flow from product discovery to implementation, testing, and release.
But autonomy also introduces a new engineering problem: someone still has to be accountable for what the agent does.
That point appeared repeatedly throughout the discussion.
Faster Building Makes Validation More Important
The panel also highlighted an interesting inversion in product development.
Producing a prototype is becoming dramatically easier.
Research, competitive analysis, interface generation, coding, documentation, and early experimentation can all be accelerated by AI.
But faster generation does not automatically create a production-ready product.
Security still needs verification. Architecture still needs review. Requirements still need validation. Generated code still needs testing. Product-market assumptions still need real users.
As one discussion thread suggested, the bottleneck may gradually move from building software toward judging whether the software should be released.
That changes what companies should look for in an AI product engineering partner.
Raw code-generation speed becomes less interesting than the combination of:
AI acceleration + engineering judgment + governance + production accountability.
With that criterion in mind, these five companies stand out.
1. GeekyAnts
Best fit: Companies building AI-native products or adding agents to existing digital products
GeekyAnts is an interesting company to include because the panel itself offers a glimpse into how its leadership is thinking about changing development teams rather than simply adopting another coding assistant.
The company has since documented an Agentic Development Life Cycle, where agents support areas such as planning, implementation, testing, documentation, and analysis while engineers retain responsibility for architecture, security, quality, and release decisions.
That human-accountability layer is important.
Agentic development becomes risky when teams assume that faster code automatically means production-ready code.
GeekyAnts appears particularly relevant for organizations that need hands-on product engineering across AI, frontend, backend, mobile, QA, and product design rather than AI strategy alone.
Its positioning is narrower than major consultancies such as IBM or EPAM, which could make it suitable for companies looking for a focused product engineering team.
2. Thoughtworks
Best fit: Enterprises rethinking the broader software development lifecycle
Thoughtworks has moved beyond treating generative AI purely as developer assistance.
Its AI/works platform uses coordinated agents across areas including requirements, specification development, software generation, testing, modernization, and runtime operations.
What makes Thoughtworks relevant to this discussion is its focus on the system around AI-generated software.
The company has been emphasizing governance, observability, enterprise context, and specification-driven development rather than relying solely on prompting.
That connects directly with one theme from the panel: traditional Agile terminology may evolve, but the underlying principles of adapting, validating, and reducing risk will remain useful.
Thoughtworks is therefore worth considering for organizations where agentic development requires organizational and architectural change alongside new tooling.
3. EPAM
Best fit: Large engineering organizations introducing agents across multiple SDLC stages
EPAM describes its approach as AI-native engineering, with AI agents and generative AI integrated into software development processes rather than operating as isolated tools.
A particularly useful example is its work with PostNL, where an initial proof of concept reportedly expanded into more than 20 types of AI agents supporting multiple teams and business units.
Use cases included automated test generation and connecting new functionality to thousands of existing test cases.
That is closer to what agentic development may look like at scale.
Instead of asking a developer to occasionally use an AI assistant, organizations begin redesigning workflows around humans and agents working together.
For large enterprises with complicated engineering environments and integration requirements, EPAM's scale and systems-engineering background make it a company worth evaluating.
4. IBM
Best fit: Enterprise agentic development, modernization, governance, and legacy environments
IBM is approaching the same transition from an enterprise software perspective.
IBM Bob is designed as an agentic development platform capable of working across planning, execution, validation, modernization, and governance.
One particularly relevant observation from IBM's recent research is that AI-generated code can move the bottleneck elsewhere.
IBM reported that 85% of surveyed DevSecOps professionals agreed that AI had shifted the bottleneck from writing code toward reviewing and validating it.
That strongly mirrors the challenge discussed during the GeekyAnts panel.
The more efficiently machines create software, the more valuable architecture decisions, security review, domain knowledge, and controlled validation can become.
IBM is therefore likely to make more sense for large organizations that need agentic development alongside legacy modernization and enterprise governance rather than only rapid greenfield product development.
5. Globant
Best fit: Companies applying agents across product, design, coding, and testing
Globant has taken a fairly direct approach to agentic software development with its CODA suite.
Rather than limiting AI agents to coding, Globant has developed agent capabilities around product definition, application design, backend prototyping, code fixing, and testing.
That broader approach matters because one of the strongest ideas from the panel was that the boundaries separating different product functions may become thinner.
If product managers can prototype, designers can interact more directly with implementation tools, developers can generate larger amounts of code, and QA agents can continuously evaluate changes, traditional handoffs start looking different.
The challenge then shifts toward coordinating those capabilities without losing engineering discipline.
Globant is worth watching because its agent strategy explicitly spans several of these functions instead of treating agentic AI purely as developer productivity software.
Product Teams May Become Smaller, But More Accountable
One of the more nuanced predictions from the panel was that teams may initially move faster with roughly the same people, and become smaller later as organizations understand which workflows AI can reliably absorb.
That seems more plausible than the simplistic claim that AI will remove half of every engineering organization overnight.
Some responsibilities will decline.
Other responsibilities will become more important.
A future product pod might need fewer people manually producing artifacts but more people responsible for:
- agent behavior
- exceptions
- architecture
- security
- evaluation
- product judgment
- governance
The panel described this emerging model as smaller teams with heavier emphasis on judgment and accountability.
That distinction matters.
Agile Probably Doesn't Disappear. Its Mechanics Might.
The panel also raised an interesting point about Agile and Scrum.
The mindset behind Agile can remain relevant even if its current rituals become less important.
Traditional sprints were partly designed around limited human throughput. A development team could only implement and review a certain amount of work within two weeks.
Agents change that constraint.
If a large amount of implementation can happen quickly, the sprint may increasingly revolve around:
What can the team safely understand, verify, evaluate, and release?
rather than:
How much code can the team produce?
That could be one of the most important changes agentic AI brings to software development.
The Real Advantage Is Not Faster Code
The temptation with agentic AI is to measure progress using development speed.
That is probably the wrong metric.
Code will continue becoming cheaper to generate.
Prototypes will become easier to create.
The harder questions will be whether teams can determine what should be built, verify what agents create, understand system-wide consequences, protect production environments, and remain accountable when autonomous workflows fail.
GeekyAnts, Thoughtworks, EPAM, IBM, and Globant are approaching this transition from different angles.
GeekyAnts is particularly oriented toward AI-powered digital product engineering and human-led delivery. Thoughtworks is rethinking the agentic SDLC and specifications. EPAM is applying agents across enterprise engineering workflows. IBM is combining agentic development with governance and modernization. Globant is deploying agents across product, design, development, and QA.
The winner in the agentic AI era may not be the team with the most agents.
It may be the team that becomes best at deciding what agents should do, what humans should review, and where humans must remain accountable.
Top comments (0)