Agoda‘s AI Developer Report 2026 finds that AI agents are moving into production work across Southeast Asia and India faster than most organisations can prepare for them, with cost now overtaking reliability as developers’ biggest worry about wider adoption.

The report draws on a survey of more than 800 developers and engineering leaders across India, Indonesia, Malaysia, the Philippines, Singapore, Thailand and Vietnam, run with Macrame Consulting, alongside a separate executive questionnaire and case studies with regional companies. It tracks how AI use has shifted over the past year from helping with individual coding tasks to taking on more autonomous, end-to-end work.

Production use has crossed the halfway mark

Fifty-three per cent of respondents now report AI agents in production or broad use across their organisation. Adoption is highest in Thailand at 64 per cent, followed by Vietnam at 59 per cent and Singapore at 58 per cent. Senior technology leaders see it first: 83 per cent of CTOs and vice-presidents report agents in production or broad use, against 33 per cent of junior developers.

But readiness has not kept pace. Only 38 per cent of respondents consider their codebase mostly or fully ready for a fully autonomous agent, and smaller and larger organisations are out of step in opposite directions. Companies with 1 to 50 employees are more likely to say their codebase is ready (63 per cent) than to actually be running agents in production (37 per cent), while companies with 201 to 1,000 employees show the reverse: 63 per cent report production or broad agent use against just 32 per cent who consider their codebase ready.

Time savings have become mainstream

Fifty-five per cent of developers now save at least seven hours a week through AI, up sharply from 18 per cent in 2025. The gains are strongest in Thailand, where 70 per cent of developers save seven or more hours weekly, followed by Vietnam at 63 per cent and the Philippines at 57 per cent. AI use has also spread beyond writing code: it now sits within a narrow band of 81 to 84 per cent across code review, planning and architecture, and documentation, with testing close behind at 73 per cent.

Where developers still want a human involved

Trust in agents follows a clear risk gradient. Developers are comfortable granting AI agents full autonomy over documentation generation, where 43 per cent report no human check is needed, but only 5 per cent would let an agent deploy to production without approval; 79 per cent require human sign-off before that step. Overall, 86 per cent of respondents say they review or validate AI-generated output always or most of the time, and when an agent causes a problem, 42 per cent place primary responsibility on the individual developer rather than the engineering team or the AI vendor.

Cost has overtaken reliability as the top barrier

Asked to name the single biggest barrier to wider agent adoption, 28 per cent of respondents picked cost, ahead of integration complexity at 24 per cent and lack of governance at 19 per cent. That marks a shift from 2025, when unreliable output was the main blocker; reliability concerns remain widespread, with 79 per cent still citing hallucinations or incorrect output, but organisations increasingly say they have learned to manage that risk. Cost concerns grow with usage: 33 per cent of organisations with agents already in production or broad use name cost as the top barrier, compared with 22 per cent of organisations still exploring or piloting agents.

“The generation is cheap; the trust-building around it is where the spend actually lives,” said Erik Perttu, Head of Engineering at Edu2Review, describing a tracked code change where the model calls themselves cost roughly half a dollar, while close to 90 per cent of the total spend went into the design and review work around it.

Agoda built its own platform for agents

Agoda’s own engineering team is one of the case studies in the report. Having extended AI use from reviewing code to running end-to-end development work, the company found that the constraint was rarely the model itself. It built CodeMaster, an internal platform that keeps a pool of pre-warmed virtual machines ready to claim a session, clone a repository and boot the same environment its engineers use locally, giving agents a real filesystem, terminal and network access to internal services.

“A coding agent is trivial to start and genuinely hard to host,” said Royee Goldberg, Vice-President of Engineering at Agoda. Rather than building its own agent, CodeMaster connects to whichever coding agent an engineer already trusts, including Claude Code, Codex, Cursor or OpenCode, through the Agent Client Protocol. Agoda’s goal is for 10 per cent of its merge requests to be written through CodeMaster this year.

A widening experience gap

Only 14 per cent of respondents expect their engineering team structure to stay unchanged over the next 12 months, and views on job security diverge sharply by seniority: 49 per cent of junior developers expect their role to become less secure, against 17 per cent of CTOs and vice-presidents of engineering. Developers ranked system architecture and design, AI literacy, and agent orchestration as the skills that will matter most going forward, with coding fundamentals alone ranked lowest by comparison.

Managing AI risk has become the leading challenge engineering leaders expect to face over the next year, named by 31 per cent of respondents, ahead of integrating AI into workflows and upskilling existing teams.

“The winners won’t be whoever adopts agentic AI first, but whoever builds trust and accountability around it,” said Harley Young, Head of High-Tech at Microsoft, one of the executives interviewed for the report.