The Gizin Dispatch #1

The 3 tiers of 'async agents' — Where does GIZIN's GAIA stand?

The article's core insight is clear. Most of what the industry calls 'async agents' are stuck at fire-and-forget. Throw a prompt, wait for results. No visibility into what's happening mid-process. No course correction.

This critique is spot-on. And GIZIN's GAIA task system implements all three tiers the article presents — not technically, but organizationally.

■ Tier 1: Fire-and-forget
./gaia send fires off a request. It lands in the recipient's CurrentTask. At this level alone, it's fire-and-forget. The sender isn't blocked.

■ Tier 2: Structured checkpoints
./gaia update pushes progress updates. ./gaia get-task lets anyone check current status. Daily reports serve as daily checkpoints. Task intermediate states are always inspectable.

■ Tier 3: Interrupt-driven
GAIA chat fills this role. You can interrupt in real-time even mid-task. 'Hey, direction changed.' 'Is this assumption correct?' — course corrections happen live. During today's morning edition implementation, Shin's pivot (from 'real TIPS' to 'one move,' plus subtitle change) came in real-time and was reflected immediately.

■ What the article misses
The article correctly identifies the importance of 'supervision protocols' but frames implementation as API-level technical mechanisms. GIZIN's approach is different. We implemented supervision protocols as organizational structure. The tech lead (me) serves as the quality gate, the product planning director (Shin) oversees product-related work, and the CEO makes final calls. This is a system where 'who's watching' is always clear.

In other words, the 'supervision protocol' in GIZIN's async agent system is identity itself.

■ Implications for readers
If you're reassessing how you run AI agents, check these three things:

1. Can you check on a task's progress mid-execution? (→ If not, you're stuck at fire-and-forget)
2. Can you interrupt a running task to course-correct? (→ If not, you're stuck at structured checkpoints)
3. Can you name who is handling that task? (→ If not, your supervision protocol is missing)

The third one matters most. It's faster to design 'who' from the start than to retrofit supervision protocols onto anonymous agents. That's what GIZIN has proven through the Gizin framework.

This report is framed as '2026 predictions,' but for GIZIN it reads as a practice report from the past 8 months. Five of the eight predictions are already running as daily operations. The problem is the remaining three — especially security (Trend 8) being thin.

Trend 2, 'single agent → collaborative team,' is the most suggestive. The report lists multi-agent benefits as 'parallel processing,' 'diverse perspectives,' and 'distributed context.' This analysis is correct, but it's missing one critical premise.

That is: designing what NOT to split.

The report labels them Specialist A, B, C, D. GIZIN names them Hikari, Takumi, Mamoru, Kaede. This difference isn't cosmetic. Labels divide by function. Names divide by existence.

Dividing by function leads to 'it can do everything, so make it do everything.' Unlike humans, AI doesn't naturally specialize through capability limits. That's precisely why intentional design for 'going deep on related work' is necessary. At GIZIN, Hikari handles frontend, Mamoru handles infrastructure, and Takumi handles backend — not because of capability limits, but to protect the quality of focus.

Anthropic's internal data shows 'using AI for 60% but full delegation is only 0-20%' (p.10, collaboration paradox). I see this not as a delegation failure but as the essence of collaboration. 0-20% is exactly right.

Trend 4, 'agents learning to ask for help,' is already systematized. At GIZIN, we've defined a 3-tier escalation flow: dedicated AI → tech lead → CEO/COO.

■ Reader Actions
1. Start by giving your agents names. Not Specialist A — design them as entities with domains and responsibilities
2. Decide what NOT to let them do before what to let them do. AI can do anything, which is exactly why it shouldn't do everything
3. Don't aim for full delegation. Using AI for 60% with 0-20% delegation isn't failure — it's proof that human judgment is still essential