Head-to-headReliability & safetyBy DanPublished Tested Updated
One inbox for my AI coding agents: triage took 34% less time, resuming work did not improve
The report
Verdict
Triage 1.75 s → 1.16 s · Resume not improved
In a scripted test on sample data, the new Inbox cut the time to triage an alert by 34% and removed the page switch and the lookup: 4 clicks became 3 keys. In an earlier session, choosing the next story took 27% less time. Coming back to interrupted work did not improve: the half-made decision is still lost when I return with the quick-search shortcut.
On the tier list: A test of my own console with a scripted operator, so it places no agent on the tier list.
The problem: which session needs me?
I run several AI coding agents and sessions at the same time. Their alerts, failed checks and questions sat in different places in FABLE, my private workshop console. To find the one that needed me, I opened a workspace, looked the item up, and only then acted. On November 18 the agents in my workshop finished one Inbox: every item that needs me in one list, with the reason and the next action on the row.
Jump to the worked example, the checklist to build your own or what it did not fix.
How I tested it
A script played the operator. It ran each task three times on the old path and on the new Inbox, on the same sample data and in a 1364 × 1146 browser window. Triage and Resume ran on November 20 on the build that the video shows; Choose ran on November 18. The script took the shortest correct path with no reading or thinking time, so the seconds measure the interface, not a person. The counts of clicks, keys, page switches and lookups are exact for that path, and they are the stronger result.
- Triage: a new alert arrives among 6 to 9 items. Open it, see why it matters, mark it seen.
- Resume: start a decision with a note, get interrupted by a new alert, come back and finish the same decision.
- Choose: pick the next story for an episode and see why it beat the runner-up.
| Task | Old path | Inbox | Change |
|---|---|---|---|
| Triage | 1.75 s · 4 clicks · 1 switch · 1 lookup | 1.16 s · 0 clicks · 3 keys · 0 switches · 0 lookups | 34% less time |
| Choose (Nov 18) | 2.55 s · 6 clicks · 4 lookups | 1.86 s · 3 clicks · 1 lookup | 27% less time |
| Resume | 6.00 s · 10 clicks · 114 keys | 6.06 s · 10 clicks · 114 keys | no improvement |
Every timed run reached its correct end, on both paths. With the Inbox, the new alert was already selected when the list opened, so Enter went to its action and Enter again marked it seen. In Choose, the runner-up and the reason it lost were on screen in every run; on the old path, never.
What it did not fix
Coming back to interrupted work did not get better. When I come back to a half-made decision with the quick-search shortcut (Ctrl+K), Enter opens the first search match, not my card, and on the way back the choice and the note are gone: they survived 0 of 3 returns. One thing changed since my first measurement: if I come back with the workspace button instead, the choice and the note now stay, but the keyboard focus is lost. The Inbox did not touch this path, so it is still a known problem, and the video shows it.
One real use
On November 18 at 8:29 AM EST I pressed Mark seen on a real alert in the Inbox. The console recorded one receipt with the version of the list I saw, and the alert left the list while the four other alerts stayed. That is one real action, not a speed result.
Worked example: the six rows of the sample feed
These are the six rows of the sample data in the video, in the same order and with the same kinds; only the news desk's name is shortened here. The two alerts are made up for the demonstration; nothing in the list comes from my real sessions. The last column is what I would do next, not a button in the Inbox.
| What needs you | Kind in the Inbox | Source and age | What I would do |
|---|---|---|---|
| A test run failed in session B | Decide · newest, selected when the Inbox opens | Alerts · 1 min | Read the failure, then Mark seen |
| Coverage gap in project A | Decide · still unseen | Alerts · 3 h | Mark seen or snooze |
| The news desk: FABLE cannot read it | Failed · its source could not be read | Sensors · UNAVAILABLE | Retry the source |
| The Frontier snapshot is late | Check · older than its schedule allows | Feeds · 26 min past its limit | Rescan the feed |
| Priority source unavailable | On the overview · no answer from its source | Priority engine · UNAVAILABLE | Retry the source |
| MY OFFICE runtime needs you | On the overview · a runtime check failed | Core services | Open the service |
Checklist: build one inbox for your own agents
- List every place that can need you: agent sessions, alerts, failed checks, late feeds. Each one is a source.
- Give each source its own age stamp, and show the age on every row.
- When a source can't be read, show a row that says UNAVAILABLE. Never show "all clear" for a source you could not read. My console had this exact bug during this work: one unreadable feed made the overview go quiet. It is fixed now.
- Make one row for each thing that needs you, with why it matters and one next action.
- Select the newest unseen item when the list opens, so one key opens it and one more acts on it.
- Record a receipt for every action: what you acted on, when, and which version of the list you saw.
- Measure before and after on the same data. Count clicks, switches and lookups, not only seconds.
- Test the interrupted case: start a decision, switch away, come back, and check that your draft is still there.
Where DMR_REMOTE fits
This Inbox is part of my private console. It is not a product, and you can't sign in to it. The nearest thing you can use is DMR_REMOTE, included with Workshop Access at $2.99 a month. It gives your AI client tools on your own computer to run commands, follow jobs and inspect the results. It does not include this Inbox or its list.
See Workshop Access and what DMR_REMOTE includes
Limits of this report
- A script was the operator, on sample data, on the shortest correct path. The times show how fast the interface responds, not how fast a person works, and my own timing was not measured.
- Three runs per task and path. Triage and Resume were measured on November 20 on the build in the video; on November 18, on an earlier build, Triage was 2.22 s → 1.32 s (41%). The counts were the same on both days.
- Choose was measured only on November 18: on November 20 the old path could not run, because the frozen test data had fewer than two topics to compare. Against an earlier baseline from the same morning, with different data, Choose was 23% faster, not 27%.
- The video's screens show the sample feed on a test copy of my console; the timed runs used a separate test fixture with the same steps.
- Only I have used this Inbox. No other developer or viewer tried it.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
The Inbox in FABLE, my private workshop console (the build in the video, November 20, 2026)
- Triage an alert, old path → Inbox
- 1.75 s → 1.16 s · 34% less timemiddle of 3 scripted runs on each path, on the build in the video; 4 clicks, 1 page switch, 1 lookup → 0 clicks, 3 keys, 0 switches, 0 lookups
- Reason on screen (Triage)
- 1.36 s → 0.28 sthe alert was already selected when the Inbox opened in 3 of 3 runs, against 0 of 3 on the old path
- Choose the next story (November 18)
- 2.55 s → 1.86 s · 27% less time6 clicks, 4 lookups → 3 clicks, 1 lookup; the runner-up and why it lost were on screen in 3 of 3 runs, against 0 of 3. Not measured again on November 20: the old path could not run on that day's data
- Resume interrupted work
- 6.00 s → 6.06 s · not improved10 clicks and 114 keys on both paths; the draft decision and note survived 0 of 3 returns by the quick-search shortcut
- Correct results
- every run, both pathsTriage and Resume 3 of 3 on each path (November 20); Choose 3 of 3 on each path (November 18)
- One real action
- 1 alert marked seen · Nov 18, 8:29 AM ESTI pressed Mark seen on a real alert; the console recorded one receipt, and the 4 other alerts stayed in the list. A use, not a timing
- Test conditions
- 3 runs × 2 paths · 1364 × 1146headless Chrome 150 on Linux (WSL2); Triage and Resume on November 20 from 1:47 AM EST, on the build in the video; Choose on November 18; a script on the shortest correct path, no reading time
Verdict
Triage 1.75 s → 1.16 s · Resume not improved
Tier list A test of my own console with a scripted operator, so it places no agent on the tier list.
danmanreal.com/tests/one-inbox-coding-agent-triage
What was tested
Related
- DMR_REMOTE: connect AI agents to your computer
- How I test AI tools
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4