AI agents are moving software work from code generation into planning, review, deployment, operations, and long-running autonomous execution.
2
Agents should be isolated from credentials and sensitive systems, with sandboxes, proxies, deterministic controls, logging, and human approval for risky actions.
3
As software becomes cheaper to produce, judgment, taste, evaluation, ownership, local context, and the design of reliable agent environments become more valuable.
Summary
Day 1 of AI Engineer Singapore brought together talks on deploying agents in software, government, design, robotics, voice, and enterprise systems. Vivian Balakrishnan described a personal second-brain workflow built from NanoClaw, WhatsApp, graph memory, local models, Whisper, and Obsidian. He argued that people can outsource computation and memory, but not personal understanding or accountability. Several speakers focused on safety. NanoClaw, Daytona, Sonar, and Simular described isolation, credential proxies, network controls, approval layers, and guardrails that sit outside the planning agent. Other talks examined what changes when code is cheap: Vercel stressed documentation and ownership, Greptile showed that agents produce different bug profiles from humans, and Stripe described one-shot coding agents that create pull requests through judge and diagnostic loops. The afternoon expanded into multimodal collaboration, world models, robotics, long-horizon reasoning, sovereign AI, voice reliability, and enterprise deployment. The common concern was practical control over increasingly capable systems.
Personal understanding and accountability remain human responsibilities
Vivian Balakrishnan argued that AI can take over calculations, memory, replication, and dissemination, but people cannot outsource personal understanding. Someone in authority can delegate work, he said, but cannot delegate accountability. He described a personal agent built around NanoClaw, WhatsApp, graph memory, semantic search, Whisper, and Obsidian. It collects speeches, transcripts, and other material, then helps with travel, meetings, briefs, and speeches. The system runs on a Raspberry Pi with 8 GB of RAM. Balakrishnan presented this as tool assembly rather than traditional software development, while still checking the code and approving sensitive access.
Agents need isolation because instructions cannot provide security
Gavriel Cohen described NanoClaw's agent factory, which reviews pull requests, creates testing plans, launches virtual machines, and runs tests before code is merged. He said pull requests are unsanitized input and that instructions such as 'never run drop database production' are not security controls. NanoClaw treats agents as if they operate behind enemy lines. Each agent runs in its own container, while credentials stay outside the agent environment. A vault proxies requests and inserts credentials only when policy allows it. Sensitive actions happen outside the sandbox after a human approves them, so the agent proposes a merge but does not hold the credentials that execute it.
The bottleneck in software is moving beyond code generation
Thibault Sottiaux described software delivery as a pipeline whose build stage has widened because agents can generate code quickly. The constraints then move to planning, review, validation, CI, security, release, operations, debugging, and understanding what happened. Codex is positioned as an agent that works across those stages. Sottiaux also discussed approval fatigue. Codex's auto-review system uses a second agent to compare actions with the original task intent, blocking suspicious or high-risk actions and redirecting the main agent. He said this reduced approvals by a factor of 20 inside OpenAI.
Agent-native systems require explicit documentation and careful ownership
Jimmy Lai said software frameworks now have to serve agents as well as people. Agents read documentation literally, copy examples, run commands, and follow errors exactly. A misleading document can confuse many projects before anyone notices, while a vague error can make an agent keep retrying and spend money. He said teams should define what fast, reliable, and secure mean in their own systems so agents can check those properties. AI makes creation cheap, but ownership remains expensive. Forking a framework means owning its security, reliability, and long-term maintenance, so teams should be careful about what they decide to build and support.
AI-generated code has different bug patterns from human code
Greptile analyzed more than five million pull requests and found strong evidence that 27.6% of April pull requests were fully vibe coded. The speakers compared revert rates, review severity, and the number of review rounds across agents and human authors. Some agents had lower revert rates than the human baseline, and some reached merged code in fewer review rounds. The results did not produce one overall winner because the outcome changed with the metric. The more useful finding was that agents produce different bugs. Cursor background agents were more likely to create N+1 query errors, while Claude agents were more likely to miss tenant checks. Validation systems need to account for those different profiles.
Long-running agents need checkpoints, memory, and the ability to change direction
Z.ai's Zixuan Li and Prime Intellect's Jackman Ong both focused on long-horizon work. Li said long horizon means depth and the ability to keep finding meaningful improvements, rather than simply running for more hours. Models can forget the original goal, accumulate errors, or continue pushing a bad approach without pivoting. He recommended checklists and repeated goal reviews, with checkpoints for verification. Ong argued that recursive language models can keep large context outside the prompt and use code structures such as loops, variables, and subagents. Training models on these scaffolds may make long-running work more reliable than relying on increasingly large context windows.
Creative and consumer AI should preserve exploration instead of forcing quick answers
Annie Luo argued that fashion and travel decisions are subjective, so efficiency is not always the right product goal. People often discover what they want by seeing several options and comparing them. Google's virtual try-on lets users see themselves in different clothes, while Google Travel uses maps as a place to wander rather than only as a destination picker. These interfaces help people build taste and confidence through the process. Phil Hedayatnia made a related argument about design agents. Training on finished designs does not teach the reasoning behind them. AirFoil's Melt stores references, metadata, comments, and annotations so models can use a designer's intent rather than copying surface patterns.
Across the talks, speakers described systems around the model that determine whether agents are useful. Stripe's Minions create pull requests from Slack prompts, run tests, send results to an LLM judge, and use a diagnostic agent to continue work when the task is incomplete. Simular separates planning from guardrails and combines accessibility trees with visual grounding for computer use. Groq described routing requests across regions and model instances to reduce latency. Swyx framed this as the agent platform above the model and container, including identity, stateful sessions, tools, permissions, model diversity, enterprise controls, and evaluations. The practical work is in making agents observable, bounded, and dependable.
Local deployment needs local data, evaluation, and control
Stefania Druga described sovereign AI as local agency over global capability. Her stack includes data, evaluation, post-training, routing, user interaction, and governance, with physical choices about on-premise and cloud compute. Sakana AI's Japanese products use local data, dialects, formal language registers, and routing to specialized or secure models. Its switchboard can send Japanese-context requests to a post-trained Japanese model, sensitive requests to an on-premise model, or requests to human review. Druga argued that different countries should choose which layers they own. Sovereignty is therefore a set of practical control points, rather than only the creation of a national foundation model.
"AI made creation really cheap, but ownership is much more expensive than you think it is."2:58:49
Who should watch
You are building coding or computer-use agents and need practical patterns for isolation, credential handling, approval flows, and testing.
Your team is producing more AI-generated code than human reviewers can comfortably inspect, and you need better validation, evaluation, and ownership rules.
You are designing consumer, creative, voice, robotics, or government AI products where local context, human judgment, exploration, and trust matter as much as raw model capability.
