Managing Agentic AI at a Development Hub

What Agentic AI Revealed After It Solved Code Generation — An AI Management Guide for Japanese Companies with Development Hubs in the Philippines

For Japanese companies with development hubs in the Philippines, this guide explains in practical terms how to manage the growing volume of AI-generated code and prevent runaway costs and falling quality.

Part 1: Why This Matters

Step 1: The Philippine Business Context (3 min)

The Philippines is known as a country rich in English-speaking talent, where software development and BPO (hubs that take on outsourced work from overseas) are concentrated. Many Japanese companies entrust system development and call-center operations to their Philippine hubs.

This article deals with how agentic AI — AI that carries out work autonomously in place of a person — sped up the work of writing code all at once, while bringing a different problem to the surface. Even when you can produce code faster and in greater volume, product quality does not improve at the same pace. That is because the truly hard parts are deciding the right requirements, connecting complex systems together, and operating them safely over the long term.

For Japanese companies with development teams in the Philippines, this is a close-to-home challenge. When the local team produces large volumes of code with AI, the burden on the people who review it suddenly increases. If review can't keep up, it leads to outages and unexpected costs. That is exactly why you need to think now about how to manage AI.

Monday morning, an office in Manila. A Japanese manager turns to the local IT head and says: "The amount of code the dev team produced with AI last week was incredible. But it looks like review isn't keeping up. I read something on this over the weekend, and it left me uneasy about chasing speed on its own. I'd like to think through how to manage this together — is that all right?" The IT head nods and peers at the screen alongside the manager.

Step 2: Key Points from the Original Article (5 min)

We have summarized the facts and arguments from the original article into a table for study purposes.

Point	What the original article presented
Only code generation got faster	What agentic AI shortens is merely the effort of implementation itself; the genuinely hard parts — judging what should be built, where responsibility lies when something breaks, and the effort of operating it once live — remain as they were
The new bottleneck is human review	The more code AI produces, the more human review backs up, and the ability to catch mistakes declines
A real case of runaway costs (Uber)	Uber used up its 2026 budget by April and put a ceiling on its AI spending
A $500 million bill in one month	According to Axios, one company's AI processing repeated endlessly, and it was billed $500 million by Anthropic in a single month
Three stages to proceed	It is recommended to put things in order in this order: (1) cost and crisis management, (2) technical policy, and (3) people and organization
Give AI minimal permissions	Don't hand AI the same strong permissions as a human as-is; require human approval for operations that change the production environment
Don't rely on one vendor or one model	Because strengths differ by product, use multiple providers and multiple models, choosing among them as appropriate
Spend on higher-performing models	Rather than a cheaper model, a high-performing model that reduces rework ends up being the better deal
Change what you measure	Measure not by lines of code but by metrics that tie to business outcomes, such as feature uptake and outage rates
Don't cut staff in a hurry	The article warns that cutting headcount before actually validating AI's results in production is premature, and that it is a decision made without correctly grasping the staffing level your company needs
Source: VentureBeat — "Agentic AI solved coding — and exposed every other problem in software engineering" (June 7, 2026)
This table was created for learning purposes based on facts from publicly available information. For details, please check the original article at the link above.

Step 3: Comprehension Check (5 min)

Q1. According to the original article, what are the things agentic AI cannot shorten? Try naming three.

Hint: Work time shrinks, but ambiguity in requirements, responsibility, and operational complexity are a separate matter.

Q2. With more code produced by AI, what new bottleneck (a place where work backs up) has emerged?

Hint: Focus on a certain task that people do by hand.

Q3. What action did Uber take regarding its 2026 AI budget?

Hint: Recall when it used up the budget, and what it put in place afterward.

Q4. Why must you not hand AI the same permissions as a human as-is?

Hint: Think about the relationship between permissions and "the ability to bear responsibility."

Q5. The original article says there is something you should do before cutting headcount. What is it?

Hint: The clue is "where" you should confirm AI's effect.

Part 2: Putting It Into Practice

Step 4: Adoption Steps in the Philippines (10 min)

We have summarized the steps for using agentic AI safely at a Philippine development hub into four stages.

Stage	What to do	Points specific to the Philippines
1. Set a cost ceiling	Decide a monthly usage cap and separate budgets for development and production in advance	Build the budget in pesos and leave a little room for exchange-rate swings. Tell the local team from the start that "once it's used up, it stops"
2. Restrict AI permissions	Allow AI read-only access, and require human approval for operations that change production	In line with the rules of the NPC (National Privacy Commission), which administers the Philippine data-protection law, set things up so AI cannot handle personal data freely
3. Designate the reviewers	Spell out in writing who reviews which code, and build review time into work hours	Don't leave it at a verbal agreement; always put it in writing. In the Philippines, verbal promises can later turn out to differ
4. Measure by outcomes	Evaluate by outage rates and feature uptake, not by code volume	Explain it in a way that fits the local evaluation system, and carefully convey the reason the metrics are changing

Step 5: Common Mistakes and How to Avoid Them (5 min)

Mistake 1: "Rewarding speed alone"

If you tell the local team "write lots of code and you'll be rated highly," thinly reviewed code piles up in volume.

Bad example: At the weekly meeting, the manager praises only lines of code or the number of items handled. The team focuses on increasing volume, and quality checks get pushed back.

Good example: You put outage rates and the scarcity of bugs found after the fact at the center of evaluation. You tell everyone at the very first meeting that the goal is to be "fast, and hard to break."

Mistake 2: "Leaving AI with strong permissions"

This is the mistake of granting AI permission even to change the production environment, just because it's convenient.

Bad example: A developer lets AI use their own account's permissions as-is, and the AI rewrites production data. No one knows who approved it, and responsibility becomes unclear.

Good example: You allow AI read-only access. Before any operation that changes production, you put in place a system where a human always checks the contents and approves.

Mistake 3: "Cutting staff before confirming the effect"

This is the mistake of immediately cutting the local team's headcount just because AI made things faster.

Bad example: Management decides "with AI, half the staff is enough," and cuts even the people responsible for review and design. As a result, outages increase and the team can't keep up.

Good example: First, you measure AI's effect in production for several months and confirm which tasks require human hands. Then, if needed, you reshuffle roles.

Part 3: Going Deeper

Agentic AI is AI that, without being instructed one step at a time, works out the procedure on its own and executes it toward a goal. At Philippine development hubs, it is becoming common to hand this AI work like test automation and code drafting, freeing local engineers to spend their time on design and review.

Least privilege is the basic rule of granting people and AI only the minimum permissions needed for the job at hand. At Manila development teams, allowing AI only to read data and having humans approve production changes helps prevent unexpected accidents.

Human-in-the-loop is a setup that inserts human checks partway through AI's work, with a human making the final call on important operations. At Philippine hubs, the practice of having a local lead always look over AI-written code before it goes into production is an example of this.

Multi-model strategy is the approach of not relying on a single AI but choosing the AI best suited to each task. For example, if a Manila team has one AI summarize text and another generate code, it can keep the impact small when one of them goes down.

Technical debt is a term that likens code written as a stopgap, which later spawns major rework, to a debt. At Philippine hubs, leaving everything to AI causes code that works on the surface but is hard to fix to pile up, becoming a cause of time and cost overruns in later maintenance.

Step 7: Applying It to Your Own Company (10 min)

Make visible how far your AI costs could balloon

Something to think about: Do you grasp how much your development team currently uses AI and how much it pays per month? As Uber did, don't you need to set a ceiling before you use up the budget?

Sort out AI permissions and "where responsibility lies"

Something to think about: How far can your AI act on its own right now? Check whether there is a system in which human approval is inserted before operations that change production.

Rethink how you evaluate the local team

Something to think about: Are you evaluating people by code volume? If you were to measure by business outcomes and the scarcity of outages, talk through which metrics would fit.

Next action: First, summarize your AI usage — monthly costs and the usage cap — on a single sheet, and share it with the relevant departments at next week's meeting.

Part 4: FAQ

Q1. Does even a small development team in the Philippines need this level of management?

Even if you are small, we recommend deciding just two things in advance: the cost ceiling and AI permissions. As in the original article, an accident in which a huge bill arrives in a single month can happen regardless of team size. A simple table is enough at first, so start with cost and permissions.

Q2. Should AI rules be unified between the Japanese head office and the Philippine hub?

The realistic approach is to align the basic policy at head office while adjusting the finer operations to fit local circumstances. The Philippines has its own data-protection law, and data handling must follow the local rules. Lay out common rules, then make clear the range in which the local side can act at its own discretion.

Q3. Even if we tell the local team "we won't cut staff," won't anxiety remain?

It is effective to convey that the change in roles brought by AI is an opportunity for growth, not a headcount cut. Show a concrete path: moving people from writing code to roles responsible for design and review. In the Philippines, showing a view of the future tends to bring reassurance and also helps improve retention.

Q4. Should AI costs be managed in pesos or in Japanese yen?

Because the actual local payments occur in pesos, showing the local team a peso-denominated ceiling makes it easier to understand. Meanwhile, compile reports to head office in yen and factor in exchange-rate swings. Managing it side by side in both currencies prevents discrepancies in understanding.

Q5. Who bears responsibility for mistakes in AI-written code?

Ultimate responsibility falls to the person who approved that code and put it into production. That is exactly why it is important to have a system that inserts a human check rather than leaving approval to AI. In the Philippines, don't leave it at a verbal agreement; keeping a written record of the approval prevents later disputes.

Tips for Making the Most of It (3 Tips)

This week, decide just two things: the "cost ceiling" and "AI permissions" Trying to put everything in order at once means you can't get started. The runaway costs and permission accidents the original article described can be greatly reduced just by nailing down these two things first. A simple table is fine — just get started.

Stop praising code volume; put the scarcity of outages at the center of evaluation Evaluate volume and thinly reviewed code piles up. Try switching the metric you highlight at the weekly meeting from lines of code to outage rates. The way the team operates will change.

Decide whether to cut staff only after measuring AI's effect in production for several months Cut headcount right away just because things got faster, and review and design will stop working. First confirm the effect in production, identify which tasks require human hands, and only then consider reshuffling roles.

Bonus: How to Make Use of PH AI Works

PH AI Works is a company that supports the use of AI and technology in the Philippines. On this article's theme — the safe adoption and management of agentic AI — we can help in a way that reflects local circumstances.

As a next step, you can consult us on things like the following:

Working together to sort out how to set cost ceilings and permissions for AI use at your Philippine hub
Supporting the building of a review system for AI-written code and the rethinking of the local team's roles
Checking how to handle AI data in line with the Philippine data-protection law

Please feel free to get in touch.

What Agentic AI Revealed After It Solved Code Generation — An AI Management Guide for Japanese Companies with Development Hubs in the Philippines

Part 1: Why This Matters

Step 1: The Philippine Business Context (3 min)