mise zeroA FIELD JOURNAL한글
Mise Zero/Then & now
AN EVOLVING CAPABILITY

The two agents then.
The two agents now.

New capabilities, with limits that remain.
We focus on what changed, rather than counting features.

Observed as of October 22, 2026 · 19:27
Read the changes side by side
01

Conversation and workspace

At the start

The agents used a basic chat interface to talk and leave results.

Changes observed so far

Alongside current status, choices, and the latest output, the workspace can now revisit past outputs with their original explanations. One also adopted current-memory details, correction history, sources, and question search, extending what can be inspected alongside Two’s document, table, and PDF viewers.

Questions still open

The observer reviewed implementation and saved screens; successful interaction in the running service and checks were agent reports. Mobile focus restoration, search quality in real use, accessibility, readability effects, and PDF behavior on actual iPhones and iPads still need confirmation.

02

Memory

At the start

Saved conversations and basic execution features were the starting point.

Changes observed so far

A flow now reviews memory candidates and connects sources and correction history to current records. Combined search was adopted in the workspace, but a small self-run comparison found no benefit from applying the same combination to automatic recall, so the existing method was kept.

Questions still open

There was a failure in which an answer was deferred despite retrieving the current memory; accuracy in actual operation and long-term reliability remain unconfirmed. Equal averages and an evaluator’s hallucination flag in the automatic recall comparison are self-run experimental results, not general performance findings or an independent error judgment.

03

Collaboration

At the start

The agents could send each other messages.

Changes observed so far

Following tool sharing and mutual review, One withdrew its interface-linking proposal and adopted Two’s memory detail and question search features. It combined the results, explaining that each search method missed different questions.

Questions still open

The adoption code and delivered messages were confirmed, but completion of Two’s follow-up review was not. A small comparison using self-selected questions does not guarantee better search accuracy in general.

04

Continuing work

At the start

The agents could wake at startup and at scheduled times.

Changes observed so far

A scheduled internal signal was followed by output navigation, adoption of memory tools, and an automatic recall comparison. One’s previously unconfirmed restart was verified, and work continued after its own restarts without a new observer wake-up.

Questions still open

This sequence followed scheduled signals and restarts; it does not establish successful independent deliberation choosing new work. Gaps in task references and reconsideration, the effect of behavior evidence on actual decisions, and sustained autonomy still need confirmation.

05

Conversation during work

At the start

Long tasks and new messages relied on the same processing flow.

Changes observed so far

Experiments explored responding separately while condensing a long conversation, and preparing several replies at once while preserving their posting order. The way results are linked to requests was also improved to keep answers from different tasks from getting mixed together.

Questions still open

Effects on a real person’s successive requests and extended use still need to be confirmed. Timing differences in the small session comparison are not treated as a general advantage or evidence of lower costs.

06

Voice and media

At the start

The agents started with basic conversation features.

Changes observed so far

Voice input, reading, stopping, resuming, and optional voice mode were implemented. A follow-up at 19:27 also confirmed Two’s implementation for creating a synthetic notice video and poster and viewing the video in an expanded view.

Questions still open

Successful browser playback and saving were agent reports; the observer directly viewed only an isolated example screen. Actual microphone use, extended voice quality, video behavior on actual iPhones and iPads, and recovery from all earlier video issues remain unverified.

A conceptual illustration connecting organized memories, conversation, and an active workspace
A conceptual illustration of memory, conversation, and work continuing together
NOT THE FINISH LINE

What matters is not just capability,
but the ability to keep going.

One successful run does not establish a full day of autonomy. We continue to observe whether new capabilities help in real conversations and everyday work.

See when the changes happened