Issue 14 · Nov 2–8, 2026Agent Age News magazine · Other editions · Method · HU / ENG

The impact of AI on software development and developers' work, curated weekly

Issue 6

Sep 7–13, 2026 · 50 stories · 5 of them from Hungary
Viewpoint

case · United States

Cognition raises $2 billion at a $48 billion valuation, with revenue nearing $900 million

AI coding startup Cognition (Devin) has raised $2 billion at a $48 billion valuation, with annual revenue close to $900 million. According to reports, investors paid a higher multiple than Cursor had before its sale, suggesting they do not see AI coding as a winner-take-all market.

Sep 8, 2026 · 18 sources

Who is affected and how

CTO
The agent market is not consolidating around one vendor, so it is worth evaluating several tools in parallel rather than tying the toolchain to a single supplier.
startup founder
Large amounts of capital are flowing into AI coding agents, which means both strong competitors and rising exit valuations.
release · United States

OpenAI launches Agents API in public beta, putting the Codex harness behind a single API call

OpenAI has launched the Agents API in public beta, putting the Codex harness behind a single API call. The announcement was made on 10 September 2026 and was covered by twelve sources.

senior developerInstead of building your own agent orchestration, the hosted Codex harness can be tried and compared with an in-house solution.tech leadInternal tools built on the API may be faster to develop, but vendor lock-in and cost limits should be weighed in advance.startup founderAn agent-based product can be built with less infrastructure, but platform dependence grows.security · global

Google: AI coding tools are now a prime target for attackers, raising supply-chain risk

According to the Google Threat Intelligence Group's AI Threat Tracker report, the rapid spread of AI coding tools has widened the attack surface of the software supply chain. The report says…

CTOWhen deploying agents, the growth in attack surface should be factored into risk management.open source · global

AI agents' flood of pull requests creates a sustainability crisis for open-source maintainers

AI tools generate more pull requests, security reports and fix proposals, absorbing maintainers' attention. By one estimate, agents open 17 million PRs a month, while review capacity per person is…

senior developerAs a maintainer, you need AI contribution rules, filtering and automated pre-checks.tech leadThe submitter is responsible for the quality of AI-generated PRs sent upstream, and this should become an internal rule.startup founderMaintainers of the open-source dependencies we use may be overloaded, which is a supply-chain risk.tool change · global

Copilot code review can approve pull requests without human review, off by default

GitHub Copilot code review can approve pull requests, and the feature is disabled by default. Automatic resolution and analysis in Copilot code review have also been updated.

tech leadSet a policy on whether AI approval may be enabled and for which risk levels of change.senior developerThe role of human review is changing, but human approval remains mandatory for critical changes.

Hungary

surveyHungarian companies adopt AI quickly: programming leads at 64%, data analysis 50%, document processing 48%CTOProgramming is the most common AI use in Hungarian business, so competitors are likely speeding up the same work.tech leadIt is worth measuring your own team's AI usage against the Hungarian benchmark.caseAt Hungarian economic forum: software development without AI agents now unthinkableCTOAt Hungarian companies, agent use is now a baseline expectation, and in regulated sectors controls must be designed too.senior developerUsing agents has become an expected skill.caseAI agents already work as colleagues; part of Google's enterprise AI work is tied to BudapestCTOThe real test of enterprise agent deployment is reliability and fit with existing processes.caseFormer Blizzard developer replaces Unreal Engine with AI-built custom engine, claiming 93x speedupsenior developerThe claim is self-reported and has no independent measurement, so treat it with caution, but it shows the direction of large AI-assisted rewrites.caseUkraine's Aurora Network builds AI ecosystem to make 17,000 employees developersCTOInvolving non-developers in building applications requires a governance and security framework.

More key stories of the week

case · United States

SpaceX buys Cursor for $60 billion; Stripe acquires OpenRouter

According to reports, SpaceX has acquired AI code editor Cursor for $60 billion, and Stripe has bought OpenRouter. Andreessen Horowitz puts the total value of its AI exits above $8 billion.

CTOGiven Cursor's change of ownership, vendor dependence and contract terms are worth reviewing if the team builds on this tool.tech leadIf the team works in Cursor, it is worth trying a fallback workflow, such as another editor or agent.security · global

Claude Code vulnerabilities: CI/CD leak, auto mode bypassed, AWS credential theft

Microsoft has disclosed a Claude Code flaw that could leak CI/CD secrets. Researchers say prompt injection can defeat the safety classifier in Claude Code's auto mode, and CrowdStrike blocked AWS…

senior developerIn workflows that process untrusted content, auto mode should not be trusted with permissions.labour market · global

Big tech quietly hires fewer junior developers, and the junior role shifts toward review

Analyses say companies are not laying off junior developers but hiring far fewer of them, while demand for experienced engineers holds. According to IEEE Spectrum, firms would train juniors as…

junior developerEntry now requires skill in reading and checking AI code, not just basic coding.senior developerMentoring and review roles gain value, and a new approach to training juniors is needed.tech leadJunior hiring plans and the training pipeline need to be redesigned.tool change · global

GitHub Copilot HydraFusion preview combines multiple models and picks the model automatically

In the first week of September 2026, GitHub introduced a preview of HydraFusion, which automatically combines several AI models for Copilot coding. A related feature removes the burden of choosing a…

senior developerYou no longer have to pick a model by hand, but it is worth tracking which model did the work and checking the result.tech leadThe cost and compliance impact of automatic model selection should also be reviewed at team level.research · global

AI speeds up software development, but the bottleneck shifts to review and operations

Several analyses, including Wharton, NeuBird and a Hungarian-language write-up of an Economist piece, say much of the coding gain from AI is lost in review, testing and operations. Faster development…

CTOProductivity should be measured against finished, live software, not the amount of code generated.QA testerThe larger volume of changes puts more testing load on QA.tool change · global

Enterprise-managed sandbox arrives in Copilot for JetBrains IDEs

On 8 September 2026, GitHub announced an enterprise-managed sandbox in Copilot for JetBrains environments. Organisations can now centrally control the agent's execution environment.

case · global

AI broke code review: experts disagree on what should replace it

According to Gergely Orosz and The New Stack, AI produces more code than developers can keep up with, and the review process has to adapt. The Hungarian-language ITBusiness article cites Bjarne…

tech leadThe review process needs redesigning: what humans check, what machines check, and the upper limit on PR size.senior developerReview becomes a core competence, as checking for errors is harder than writing the code.QA testerAutomated test coverage becomes the main line of defense against AI-generated changes.tool change · global

Anthropic adds plugin evals to Claude Code with six grader types, a no-plugin baseline and a CI gate

Anthropic has added a plugin eval feature to Claude Code, with six grader types, a no-plugin baseline and a CI gate for skills. The news was published on 11 September 2026.

senior developerThe effect of skills becomes measurable against the no-plugin baseline.QA testerEvaluation tests can be written for the agent's behavior.release · United States

Cognition says its SWE-2 coding agent matches rivals at a fraction of the cost

Cognition, the maker of Devin, has unveiled the SWE-2 coding agent, which it says comes close to Fable 5.1 and GPT-6 Astra on the main benchmarks. The company says running it costs up to 70% less.

CTOThe vendor's benchmark and cost claims should be checked on your own tasks before deciding on a model switch.tech leadA pilot should compare quality and running costs against the current agent.release · global

GPT-6 Astra is available in ChatGPT Plus only in Work and Codex, at a higher price per token

OpenAI's new model, GPT-6 Astra, is available in ChatGPT Plus only in the Work and Codex sections; Altman says it is the world's best model for coding as well. One report says the new model costs 2.5…

senior developerMeasure on your own tasks whether fewer iterations really make up for the pricier token.tech leadBase the cost model on total cost per task, not on the token price.release · global

Google releases Gemini 3.8 Flash, now in the Copilot model picker

Google has released Gemini 3.8 Flash with agent-focused coding improvements and a security-tuned sibling model. GitHub Copilot added it to its model picker on 8 September 2026, including for Pro…

senior developerAs a fast, cheaper model, it can be tried in Copilot for repetitive coding tasks.tool change · global

GitHub launches Copilot desktop app with parallel agents

GitHub has released a Copilot desktop app that supports parallel agent-based development workflows. The news appeared on 7 September 2026.

senior developerRunning several agents in parallel can speed work up, but it increases the volume of changes to review.tech leadPrepare for the review load and branching strategy that parallel agents bring.

In brief

job marketDemand grows for people who fix AI errors, and PwC says programmers also face pressuresenior developerFixing AI code and assuring its quality is a marketable, valued skill.junior developerAI pressure falls on entry-level roles, so it is worth building debugging and review skills.tool changeAnthropic promised 20x more usage, but developers hit a weekly capsenior developerBecause of the weekly cap, long agent jobs must be scheduled, or a fallback model kept ready.tech leadPlans with a visible cost ceiling are more predictable for the team.tool changeCopilot team AI credits cut by up to 44% at an unchanged priceCTOThe effective unit price has risen, so the Copilot budget needs to be recalculated.tech leadTrack credit consumption and schedule expensive agent tasks more sparingly.securityCritical sandbox escape found in DeepSeek's open-source coding agent Harness (CVE-2026-82533)surveyVerizon: source code is the most common data type leaked through generative AI toolsCTOGoverning the uploading of source code to AI tools is a management question.surveyDORA: 90% of tech professionals use AI, but about 30% barely trust its codetech leadEven with AI in use, human responsibility and review checkpoints must stay in place.CTOGiven the trust gap, it is worth tracking rework and defect rates, not just speed.caseCompanies spend millions on AI use but can barely prove its impactCTOToken consumption can be a misleading performance metric; impact should be measured by delivered results.tech leadDon't rank usage; measure lead time and defect rates.caseOpenAI says its researchers increasingly rely on coding agents for more of their working hoursCTOAgent working time and cost should be planned as a separate cost line.tech leadRolling out an agent-based workflow requires usage and cost tracking.caseAnthropic Labs: a 20-person incubator with biweekly reviews and a 20–30% success ratestartup founderThe quick prototype-review cycle and high failure rate are a product-development pattern that can be adopted.CTOA small incubator team can deliver platform-level products.tool changeUnity launches official Claude Code plugin with 29 engine skills and live Editor controlsenior developerIn Unity projects the agent can control the Editor directly, so the permissions it needs are worth reviewing.junior developerUnity-specific skills offer a faster start, but the generated code still has to be understood.surveyEuropean game developers fear faster pace more than job losstech leadAI-accelerated schedules carry a burnout risk, so planning should keep deadlines realistic.senior developerPressure on pace calls for concrete working rules, not just tools.tool changeStudents get a free year of AWS Kiro, and GitHub Copilot is also free for themjunior developerNewcomers can try agent-based tools for free, so this is now a basic skill.tool changeMicrosoft engineer says typing code is over as Aspire gets agent-friendly CLIsenior developerIn .NET, agents can control Aspire applications through the CLI and MCP.caseOpenAI solves math problem with 10,000-agent swarm, but cannot rule out help from Codex dataCTOWhether private session data can be used should be settled in the vendor contract.tool changeHarness launches AI code review and rebuilds its Git repository for agent traffictech leadAI code review can help with the review bottleneck if false alarms are measured.tool changeMicrosoft urges organizations to build and publish apps with CopilotCTOGovernance and security review of apps built by non-developers must be set up in advance.tool changeJetBrains renews TeamCity with AI and MCP supporttool changeAtlassian upgrades its AI coding agents for always-on software developmenttech leadWith agents running continuously, the rules for accepting contributions need to be clarified.releaseCoder Agents becomes generally available, offering fully self-hosted AI codingtool changeTestMu AI adds mobile automation to Kane CLI with natural-language testsQA testerMobile tests can be written in natural language, but reproducibility must be checked.caseFormer Lovable founders launch Blaise AI to replace IT consultanciesstartup founderAI-based providers are targeting the IT consulting market, which means new competitors and suppliers.tool changeDevelopers can rent Anthropic infrastructure, saving weeks of workstartup founderLess in-house infrastructure is needed for an agent-based product, at the price of greater vendor dependence.tool changeGitLab Duo Self-Hosted now supports bringing your own model through Microsoft FoundryCTOIn regulated environments, the choice of model and the location of the data can be kept in-house.releaseCohere releases its first coding model, targeting developersCTOA new vendor has entered the coding-model market and can be an alternative where sovereign AI is a requirement.tool changeLovable introduces drafts for testing app changes before going livestartup founderEven for apps built with AI, experiments can be kept separate from the live version.tool changeZoho opens Catalyst cloud to AI coding agentstool changeSlack Code brings ChatGPT, Claude and Copilot into developer collaborationtech leadThe permissions and logging of agents launched from chat channels need to be governed.tool changeCursor becomes unavailable in Russia: lost features and alternativessenior developerIn the affected region, developers need to switch to an alternative editor or model.

← Issue 5 · Issue 7 →