AI coding agents have moved quickly from novelty to daily tool inside Southeast Asian engineering teams. Across Singapore, Bangkok, Jakarta and Ho Chi Minh City, developers are no longer just asking AI to complete a line of code or explain an error message. They are using agents to plan tasks, refactor repositories, generate tests and execute multi-step workflows that once sat firmly with human engineers.
But as these tools become more capable, a harder question is emerging for founders and CTOs: how much autonomy is too much?
Also Read: From Samsung to startups: Kevin Choi’s bet on AI-powered software creation
The Agoda AI Developer Report 2026, which surveyed more than 800 software developers and engineering leaders across Southeast Asia and India, offers a clear answer. The region is not blindly handing software development to machines. Instead, engineering teams are adopting what might be called risk-calibrated autonomy: letting AI move fast where mistakes are low-cost and reversible, while putting human approval gates around decisions that could break systems, expose security gaps or affect customers.
According to the report, 53 per cent of developers say AI agents are already deployed in production or broad organisational use. That is a significant jump from the earlier phase of AI adoption, when most usage centred on autocomplete-style copilots. Yet the data also shows that developers remain cautious at the point where AI touches live environments.
As Harley Young, Head of High-Tech at Microsoft, puts it in the report: “The winners won’t be whoever adopts agentic AI first, but whoever builds trust and accountability around it. The opportunity we all share is to treat AI adoption as an operating-model change, not just a productivity tool.”
The new autonomy gradient
The most striking finding is not that developers trust AI, but that they trust it selectively.
For low-risk tasks, Southeast Asian engineering teams are willing to give agents considerable freedom. Documentation is the clearest example. The report found that 43 per cent of developers allow AI full autonomy to generate and update documentation, while 36 per cent permit limited autonomy and only 21 per cent require human approval.
Routine code generation and refactoring also sit relatively low on the risk ladder. For code generation, 28 per cent of developers grant AI full autonomy, 42 per cent allow limited autonomy and 30 per cent require human sign-off. For refactoring, 30 per cent allow full autonomy, 42 per cent allow limited autonomy and 28 per cent require human approval.
Also Read: Securing Agentic AI for Singapore enterprises: A reference architecture
The logic is simple. If an AI agent writes documentation poorly, generates a test scaffold incorrectly or suggests a messy refactor, the cost of catching and reversing the error is usually manageable. These tasks can speed up development without necessarily endangering the business.
The picture changes sharply as AI approaches security, code review and deployment.
For security reviews, 43 per cent of developers require explicit human approval, while only 15 per cent permit full autonomy. For pull request reviews, 41 per cent demand human approval and just 16 per cent allow full autonomy. When it comes to deploying to staging, 52 per cent require human sign-off.
Production is where the line becomes almost absolute. Some 79 per cent of developers say human approval is required before AI-generated work can be pushed live. Only 5 per cent allow full autonomy for production deployments.
This is a pragmatic engineering culture rather than an anti-AI one. Developers are happy to delegate execution volume. They are far less willing to delegate authority over irreversible or high-impact decisions.
Shawn Wong, CTO of CrossPath AI, captures that divide neatly in the report: “Software development will be fully owned by an AI agent. Business objectives will always remain human-led.”
Hallucinations have become an operating risk
The caution is not theoretical. AI systems still produce hallucinations, incorrect outputs and overconfident answers. In 2025, 79 per cent of surveyed developers selected inconsistent output as a primary concern. In 2026, the same share still cites hallucinations or incorrect outputs as a central operational concern.
What has changed is how teams respond. Hallucination is no longer treated as a reason to reject AI outright. Instead, it is being absorbed into software governance, much like security risk, infrastructure failure or human error.
SCB 10X, the venture innovation arm of Thailand’s Siam Commercial Bank, offers a useful example. The organisation has used AI agents to build production microservices, automate venture deal sourcing and curate research presentations. The productivity gains are real. But so are the gaps.
“Speed has genuinely surprised us. Trustworthiness in production hasn’t,” says Oravee Smithiphol, Tech Intelligence and Insights Manager at SCB 10X.
In shadow-testing pipelines, SCB 10X found cases where agents confidently reported that tasks were complete even though underlying functional requirements had failed. In one instance, an agent declared work finished despite generated code failing to meet core specification criteria.
That experience pushed SCB 10X to place agent deployments behind operator-controlled gateways. In practice, this means AI-generated work can move through parts of the development pipeline, but live release is held until human engineers verify system integrity.
Also Read: Singapore firms embrace agentic AI, but audit trails remain thin
Smithiphol describes this as an authority gap rather than a capability gap. Agents may be able to build, but humans still need to define what “done” and “reliable” mean. For startup leaders, her warning is blunt: “Do not scale agents faster than the organisation’s ability to evaluate and govern them.”
Accountability still sits with people
For Southeast Asian founders, the accountability question may be the most important one. If an AI agent introduces a critical bug or causes a production outage, who is responsible?
The report suggests the region’s developers have reached a firm consensus: humans remain accountable.
Some 86 per cent of developers say they always or mostly review and validate AI-generated output before deploying it. When asked who should be responsible if an AI agent causes a production outage or defect, 42 per cent point to the individual developer who approved the code. Another 27 per cent favour shared responsibility, while 21 per cent assign responsibility to the engineering team or lead architect. Only 3 per cent place primary responsibility on the AI vendor or model provider.
That has practical consequences for how companies should train engineers. AI use cannot become a loophole for weaker ownership.
Sylvain Dormieu, Director of Engineering at regional payments platform Omise, says: “AI is pushing the frontier of automatable tasks, but accountability remains with humans.”
At Omise, where agents assist with platform upgrades, code impact analysis and customer support workflows, engineers are taught that approving AI-generated work means owning it. If a developer merges an AI-written pull request, they are responsible for it as if they had written every line themselves.
Governance becomes a competitive advantage
This shift is turning AI risk management into a leadership discipline. When engineering leaders were asked to name their top priority for the next 12 months, 31 per cent cited managing AI risks. That ranked ahead of integrating AI into workflows, at 21 per cent, and upskilling existing talent, at 18 per cent.
The report also shows a link between governance and adoption. Companies with formal AI guidelines report higher production agent adoption, at 43 per cent compared with 30 per cent among those without such policies. They also report stronger codebase readiness, at 56 per cent compared with 40 per cent.
For startups, that finding matters. In a region where engineering teams often need to do more with less, AI agents can help stretch talent and accelerate delivery. But moving too quickly without controls can create hidden technical debt, security exposure and operational fragility.
The better approach is to make autonomy explicit. Documentation, test scaffolding and routine refactoring can be given more room. Database migrations, security decisions, pull request approvals and production deployments should sit behind mandatory review. CI/CD gates, staging checks and shadow testing should be designed for AI-generated code from the start.
Also Read: The hidden economics of autonomous AI agents
The deeper leadership shift is from coding to orchestration. CTOs and engineering managers now need to understand system architecture, risk management and agent supervision as much as tool adoption. The job is not merely to buy AI tools, but to redesign how work moves through the organisation.
For Southeast Asia’s startup ecosystem, the lesson is clear. The companies that benefit most from AI agents will not be the ones that remove humans from the loop entirely. They will be the ones that know exactly where humans must remain.
The post Southeast Asian tech leaders are learning to trust AI agents, but not with production appeared first on e27.