현직 9년차 메타 AX 엔지니어가 경고하는 "진짜 AX"

Main ideas / concepts conveyed

• “AX” (AI Transformation / Agent-centric Transformation) is not just adopting AI

  • In Korea, “AX” is widely discussed, but the speaker argues many people interpret it narrowly as “use AI” or “become AI-native.”
  • The speaker’s core definition: AX is an agent-centric workflow and systemic redesign—transforming how work is organized and executed when work shifts from human-centered to agent-centered.

• What AX is really about

  • Rebuilding organizational processes and workflows so tasks can be performed efficiently by agents.
  • Moving from “AI first” rhetoric (“AI is good, let’s change”) to designing work so agents can handle it.
  • The end goal is a “Human Unlock” system (speaker’s phrasing): humans intervene only when absolutely necessary.

• Terminology differences (AX vs other buzzwords)

  • The speaker downplays the importance of label differences (AX vs other Korean/English terms).
  • What matters is the underlying shift to agent-centric workflows and organizational/system redesign.

Methodology / practical approach (organized as steps)

1) Start with the correct question: transform work → agent-centric organization

• Treat AX as a transformation of people’s existing work into agent-centered workflows.
• Expect “everything” to be required (not a single tool):
  • Build/extend internal systems such as:
    • memory systems
    • tooling infrastructure
    • applications
  • Deploy Forward Diplomat Engineers (FDEs) as consultants to assist teams.

2) Don’t focus only on AI tools or “AX champions”—focus on process bottlenecks

• When a company says “we’ll do AX,” the speaker recommends challenging them to identify:
  • the biggest bottleneck in the work of the person/team.
• Concept of bottleneck placement:
  • If a task involves 1–100 steps, find where most time/resources are consumed.
• Apply AI/agents where they reduce the most painful work:
  • Examples mentioned:
    • tedious tasks like quality inspection
    • report creation and presentation prep
    • meetings (note-taking/summarization)
    • reducing manual work that can now be automated by AI
• Leadership’s role:
  • Process/culture change is more critical than selecting an AI-savvy champion.
  • Risk-taking at leadership level is necessary.
• Concrete warning:
  • Even with AI adoption, if the organization keeps rigid legacy tools/formats (example: Hancom HWP insistence), process won’t truly change.

3) Build internal “Harnness” (tooling/toolkit) rather than relying on external tools

• The speaker argues large organizations inevitably create their own internal harness:
  • “Harness” is framed as a company-specific toolkit (agent engine, workflow glue, integrations).
• Reason:
  • External coding/agent tools (e.g., “Cloud Code/Codex” type systems) are hard to integrate with internal APIs.
• Suggested pattern (implied):
  • External plugins may work temporarily, but companies eventually want deeper control → build/extend your own harness.

4) Measure success using business impact + resource efficiency (not token usage)

AX success metrics are described as having two core components:

• Resource management / cost efficiency
• Driving impact
  • Primary success idea:
    • Maximum output with the same personnel/resources
• Avoid a misleading metric:
  • “Token spending” / token matching shouldn’t be the main KPI.
  • Higher tokens do not necessarily correlate with better outcomes.

5) Use proxy metrics when immediate results don’t map directly to business impact

• If you can’t measure direct business value immediately:
  • create proxy metrics to measure efficiency improvements.
• Examples:
  • Meeting/reporting time drops from 10 hours → 1 hour (measure time saved).
  • A process took a month → final deliverable in a few days/weeks (measure cycle-time reduction).

6) Establish maintenance + feedback loops (AX doesn’t stop after rollout)

• Reject the simplistic view: “AX begins when tools are introduced and then it’s done.”
• Maintenance is part of the real transformation:
  • When systems produce wrong/strange outputs, humans should report issues into a feedback loop.
• This feedback improves agent quality continuously.

7) Evaluate models/tools carefully (tuning prompts is not enough)

Evaluation is described as one of the hardest and most time-consuming parts:

• Two evaluation types
  • Online evaluation
    • run tests using real/expected usage patterns
    • analyze results such as turn-by-turn quality and speed of producing correct output
  • Offline evaluation
    • pre-test by anticipating scenarios and running prepared cases
• Human feedback / UXR
  • incorporate feedback from people into evaluation
• Include evaluation in AX because:
  • many “AI tools” can be widely used but not actually be good.

8) When models/harness update, don’t blindly upgrade—re-evaluate

• If the underlying model or agent capability changes:
  • previously working skills/pipelines may degrade.
• Warning:
  • don’t switch everything unconditionally just because a new model is released.
  • evaluate the delta:
    • measure improvement/degradation vs previous versions.

9) Set objective quality standards and avoid role confusion / inefficiency

• Concern:
  • After AX, “vibe coding” and overlapping responsibilities can cause review bottlenecks.
• Culture implication:
  • allow more people to create output (anyone can contribute),
  • but verify objectively and scalably.
  • create standards so review/verification doesn’t become wasteful,
  • especially prevent overload when many stakeholders submit work at once.

10) Prevent “reinventing the wheel” across teams

• Large corporations can end up with:
  • many teams creating 100 similar report-writing skills/plugins.
• AX should include:
  • infrastructure/platform for sharing and verification
  • platform ideas mentioned:
    • skill marketplace management
    • quality control systems

11) Treat access rights, privacy, and security as a core AX constraint

• Security/privacy refinement is identified as one of the hardest problems:
  • even if internal agents could use conversation data, privacy comfort may become the limiting factor.
• AX requirement:
  • build infrastructure to refine/store usable data safely with robust security.
  • accept this may slow efforts, especially in large enterprises.

12) “Remove the AI feel” = iterative harness/prompt/skills engineering + evaluation

• Consistent output quality requires skill:
  • harness engineering / evaluation / prompt engineering
• Process:
  • iterative trial + feedback
  • refine based on errors or weak outputs
• Forecast:
  • human work will increasingly become feedback-driven agent training, not direct manual execution.

13) AX career path: not SI/support work only—AX is business/process core

• The speaker rejects the idea that AX is “draining” or peripheral.
• AX is described as learning to identify inefficiencies and redesign workflows/process operations.
• Claimed career outcome:
  • AX-capable people become strong business operators who can optimize processes across companies.

14) Leadership and communication are central (including English for global big tech)

• For overseas big tech:
  • English communication matters because AX is about persuasion:
    • accurately identify problems
    • propose solutions
    • align/sell to stakeholders
• The speaker downplays purely technical knowledge as the differentiator compared to persuasive/communication ability.

15) Developers’ next goal

• Same guiding KPI philosophy:
  • increase output efficiency using given resources.
• Suggested mindset:
  • take more risks
  • fail more often
  • learn and iterate
  • become an earlier adopter
  • build and persuade around solutions

16) Final framing: AX is not “click everything”

• AI doesn’t fully work by “clicking.”
• AX’s purpose:
  • refine and invest so “click-like automation” becomes genuinely high-quality and UX-focused.
• Ultimate target state:
  • an organization where agents make ~95% decisions and humans handle ~5%.
  • not necessarily meaning people get fired—work shifts to intervention only when needed.

Speakers / sources featured

• Primary speaker: A “transaction Galex” / “Tech Lead” / “9-year veteran Meta developer” (name appears garbled in subtitles as “transaction Galex” and “Alex”; exact legal name unclear due to subtitle errors).
• Organization/source referenced (not a speaker):
  • Meta
  • Agentec Transmation (speaker-referenced organization name; subtitle text is garbled)
  • KakaoTalk
  • Cloud Code / Codex (tools referenced)
  • UXR (user experience research concept referenced)