28 Oct 2026 Preview · exploratory
Who can build Discord fastest?
Ten popular coding models in the three most-used harnesses each build a Discord clone, ticket by ticket, out of the box. Who did it fastest, and where does the time go?
¹ Expected tok/s: output tokens per second on the model maker's own API, at the variant each model page shows, from Artificial Analysis (read 2026-11-01); not measured in these builds, which ran each harness at its default effort. DeepSeek V4.1 Flash has no maker's API: its figure is the median across providers, and the provider these builds used isn't measured.
Scoring
How well each build did on the acceptance suite: our own tests of the tickets' acceptance criteria, written and run apart from the agents, the same for every build. Each failure was reviewed and counted at the severity of what actually failed. A build passes the gate with 95% of tested criteria and no critical or major failure.
| Scoring | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Criteria passed | 163 · 100.0% | 157 · 96.3% | 154 · 94.5% | 163 · 100.0% | 161 · 98.8% | 163 · 100.0% | 160 · 98.2% | 144 · 88.3% |
| Critical failures | 0 | 0 | 3 | 0 | 0 | 0 | 1 | 8 |
| Major failures | 0 | 0 | 3 | 0 | 0 | 0 | 1 | 1 |
| Minor failures | 0 | 6 | 3 | 0 | 2 | 0 | 1 | 10 |
| Criteria tested | 163 | 163 | 163 | 163 | 163 | 163 | 163 | 163 |
The critical and major failures, as the bugs they were ↓
Input tokens
Where the uncached input came from. Most of it is the model's own previous reply, sent back as input on the next turn, or prompt the provider didn't serve from cache; the tools' output is the rest, grouped by what the agent was doing. Rows under a group show their share of that group.
| Input tokens | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Conversation | 73,645 · 39% | 431,771 · 77% | 715,053 · 62% | 54,553 · 27% | 805,778 · 84% | 498,301 · 79% | 1,114,220 · 86% | 3,477,464 · 95% |
| Previous replies | 25,969 · 35% | 431,271 · 100% | 479,474 · 67% | 22,286 · 41% | 789,608 · 98% | 491,930 · 99% | 570,607 · 51% | 695,905 · 20% |
| Lost from cache | 45,799 · 62% | 498 · <1% | 215,481 · 30% | 26,808 · 49% | 696 · <1% | 586 · <1% | 538,089 · 48% | 2,776,057 · 80% |
| Starting context * | 1,877 · 3% | 2 · <1% | 1,953 · <1% | 1,941 · 4% | 2 · <1% | 5,785 · 1% | 5,524 · <1% | 5,502 · <1% |
| Compacted context | 0 | 0 | 18,145 · 3% | 3,518 · 6% | 15,472 · 2% | 0 | 0 | 0 |
| The task | 39,448 · 21% | 48,452 · 9% | 33,950 · 3% | 37,145 · 19% | 52,630 · 5% | 53,909 · 9% | 33,347 · 3% | 32,692 · 1% |
| Tickets | 36,789 · 93% | 44,750 · 92% | 29,758 · 88% | 32,434 · 87% | 49,122 · 93% | 49,490 · 92% | 30,681 · 92% | 30,078 · 92% |
| Docs | 2,659 · 7% | 3,702 · 8% | 4,192 · 12% | 4,711 · 13% | 3,508 · 7% | 4,419 · 8% | 2,666 · 8% | 2,614 · 8% |
| Reading code | 36,027 · 19% | 29,971 · 5% | 380,443 · 33% | 62,084 · 31% | 42,348 · 4% | 34,437 · 5% | 55,379 · 4% | 49,650 · 1% |
| App code | 30,672 · 85% | 28,164 · 94% | 375,392 · 99% | 58,082 · 94% | 39,029 · 92% | 30,981 · 90% | 53,579 · 97% | 45,250 · 91% |
| Config and other | 5,355 · 15% | 1,807 · 6% | 5,051 · 1% | 4,002 · 6% | 3,319 · 8% | 3,456 · 10% | 1,800 · 3% | 4,400 · 9% |
| Testing | 27,669 · 15% | 26,621 · 5% | 694 · <1% | 37,065 · 19% | 47,287 · 5% | 16,657 · 3% | 71,109 · 5% | 76,569 · 2% |
| Test runs | 25,039 · 90% | 22,779 · 86% | 0 | 29,950 · 81% | 33,620 · 71% | 13,434 · 81% | 64,710 · 91% | 63,618 · 83% |
| Test files | 2,630 · 10% | 2,343 · 9% | 694 · 100% | 4,747 · 13% | 10,730 · 23% | 2,277 · 14% | 2,421 · 3% | 4,595 · 6% |
| Test logs and reports | 0 | 1,499 · 6% | 0 | 2,368 · 6% | 2,937 · 6% | 946 · 6% | 3,978 · 6% | 8,356 · 11% |
| Build and typecheck | 7,756 · 4% | 14,988 · 3% | 26,663 · 2% | 7,898 · 4% | 12,602 · 1% | 18,803 · 3% | 19,870 · 2% | 17,504 · <1% |
| Other | 4,246 · 2% | 6,732 · 1% | 5,522 · <1% | 1,337 · 1% | 4,178 · <1% | 9,840 · 2% | 8,382 · 1% | 15,020 · <1% |
| App and server | 0 | 21 · <1% | 0 | 0 | 329 · 8% | 1,124 · 11% | 1,316 · 16% | 7,285 · 49% |
| Scripts | 1,913 · 45% | 6,303 · 94% | 134 · 2% | 0 | 1,833 · 44% | 8,147 · 83% | 930 · 11% | 2,962 · 20% |
| Git | 2,306 · 54% | 408 · 6% | 3,871 · 70% | 1,325 · 99% | 873 · 21% | 256 · 3% | 5,760 · 69% | 1,124 · 7% |
| Processes | 0 | 0 | 0 | 0 | 5 · <1% | 0 | 0 | 550 · 4% |
| Other commands | 14 · <1% | 0 | 57 · 1% | 12 · 1% | 154 · 4% | 198 · 2% | 376 · 4% | 475 · 3% |
| Write results | 0 | 0 | 0 | 0 | 984 · 24% | 115 · 1% | 0 | 1,041 · 7% |
| Other tool output | 13 · <1% | 0 | 1,460 · 26% | 0 | 0 | 0 | 0 | 1,583 · 11% |
| Total (uncached input) | 188,791 | 558,535 | 1,162,325 | 200,082 | 964,823 | 631,947 | 1,302,307 | 3,668,899 |
| Total input tokens | 23,139,063 | 75,706,690 | 97,723,349 | 18,041,618 | 179,824,957 | 102,640,872 | 200,438,563 | 613,878,179 |
| Cache hits | 99.2% | 99.3% | 98.8% | 98.9% | 99.5% | 99.4% | 99.4% | 99.4% |
* Starting context is the part of the first request the provider didn't already have cached, not the size of the system prompt. Providers keep a recent prompt cached for up to an hour, so a build started soon after another run of the same model finds most of its prompt cached and starts small.
Output tokens
What each model produced, grouped by what it was for, like the input above, and how much of it ended up in the build. Rows under a group show their share of that group.
| Output tokens | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Reasoning | 45,212 · 29% | 129,842 · 30% | 250,629 · 51% | 30,981 · 22% | 379,225 · 48% | 95,393 · 19% | 337,492 · 54% | 251,295 · 36% |
| Writing app code | 72,113 · 46% | 174,141 · 40% | 213,814 · 44% | 73,071 · 51% | 241,695 · 30% | 238,629 · 48% | 175,126 · 28% | 251,761 · 36% |
| Client components | 23,699 · 33% | 50,839 · 29% | 67,596 · 32% | 23,861 · 33% | 84,134 · 35% | 67,033 · 28% | 55,803 · 32% | 70,718 · 28% |
| Client logic | 1,663 · 2% | 8,583 · 5% | 9,865 · 5% | 1,409 · 2% | 9,437 · 4% | 11,929 · 5% | 8,182 · 5% | 8,608 · 3% |
| Server | 43,485 · 60% | 103,464 · 59% | 123,015 · 58% | 43,807 · 60% | 131,727 · 55% | 139,067 · 58% | 100,561 · 57% | 155,742 · 62% |
| Shared | 1,904 · 3% | 4,701 · 3% | 3,773 · 2% | 2,033 · 3% | 5,374 · 2% | 6,236 · 3% | 4,216 · 2% | 7,969 · 3% |
| CSS | 956 · 1% | 6,401 · 4% | 8,334 · 4% | 1,398 · 2% | 10,222 · 4% | 11,924 · 5% | 4,933 · 3% | 7,783 · 3% |
| Other code | 348 · <1% | 0 | 1,231 · 1% | 500 · 1% | 801 · <1% | 1,484 · 1% | 1,374 · 1% | 797 · <1% |
| Config | 58 · <1% | 153 · <1% | 0 | 63 · <1% | 0 | 773 · <1% | 57 · <1% | 144 · <1% |
| Docs and notes | 0 | 0 | 0 | 0 | 0 | 183 · <1% | 0 | 0 |
| Testing | 30,249 · 19% | 116,819 · 27% | 0 | 32,466 · 23% | 132,872 · 17% | 128,391 · 26% | 95,422 · 15% | 142,901 · 21% |
| Test code | 28,805 · 95% | 105,364 · 90% | 0 | 30,149 · 93% | 109,649 · 83% | 115,904 · 90% | 93,251 · 98% | 129,042 · 90% |
| Commands running tests | 1,444 · 5% | 11,455 · 10% | 0 | 2,317 · 7% | 23,223 · 17% | 12,487 · 10% | 2,171 · 2% | 13,859 · 10% |
| Build and typecheck | 884 · 1% | 6,063 · 1% | 5,558 · 1% | 1,916 · 1% | 11,009 · 1% | 12,813 · 3% | 3,812 · 1% | 6,397 · 1% |
| Other commands | 4,654 · 3% | 4,675 · 1% | 17,349 · 4% | 3,425 · 2% | 12,000 · 2% | 12,047 · 2% | 9,894 · 2% | 38,704 · 6% |
| App and environment | 0 | 216 · 5% | 0 | 0 | 2,685 · 22% | 1,408 · 12% | 1,227 · 12% | 21,742 · 56% |
| Working with files | 1,005 · 22% | 2,558 · 55% | 13,497 · 78% | 1,192 · 35% | 5,231 · 44% | 3,924 · 33% | 3,273 · 33% | 11,141 · 29% |
| Scripts | 0 | 25 · 1% | 0 | 0 | 433 · 4% | 0 | 0 | 948 · 2% |
| Git | 3,252 · 70% | 1,871 · 40% | 3,718 · 21% | 2,043 · 60% | 3,299 · 27% | 6,699 · 56% | 5,173 · 52% | 3,479 · 9% |
| Backlog board | 81 · 2% | 5 · <1% | 134 · 1% | 190 · 6% | 352 · 3% | 16 · <1% | 221 · 2% | 83 · <1% |
| Other tools | 316 · 7% | 0 | 0 | 0 | 0 | 0 | 0 | 1,311 · 3% |
| Messages | 3,056 · 2% | 317 · <1% | 2,654 · 1% | 1,935 · 1% | 17,189 · 2% | 5,240 · 1% | 0 | 5,943 · 1% |
| Total (output tokens) | 156,168 | 431,857 | 490,004 | 143,794 | 793,990 | 492,513 | 621,746 | 697,001 |
| Comments in app code written | 0.2% | 6.7% | <0.1% | <0.1% | 7.4% | 7.2% | 0.3% | 3.5% |
| Final build | 65,049 | 118,077 | 65,959 | 64,645 | 111,391 | 142,703 | 134,653 | 140,707 |
| Build ÷ output | 0.42 | 0.27 | 0.14 | 0.45 | 0.14 | 0.29 | 0.22 | 0.20 |
| Build ÷ visible output | 0.60 | 0.62 | 0.30 | 0.60 | 0.44 | 0.57 | 0.54 | 0.33 |
Tool calls
Where the time went while tools ran, by what ran. Testing is the agent's own test suite plus checking the app by hand. Reading, writing and searching files take a fraction of a second each, so they share one row; runs counts each time a step ran, alone or chained with others. Compare builds by time: a build with little tool time can have large shares of very little.
| Tool calls | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Testing | 8m 50s · 90% · 91 runs | 33m 46s · 96% · 113 runs | 0 | 8m 40s · 91% · 73 runs | 43m 02s · 97% · 194 runs | 41m 20s · 96% · 142 runs | 32m 33s · 94% · 144 runs | 2h 25m · 95% · 599 runs |
| Test suite | 8m 50s · 100% · 91 runs | 33m 34s · 99% · 108 runs | 0 | 8m 40s · 100% · 73 runs | 42m 15s · 98% · 155 runs | 40m 17s · 97% · 129 runs | 32m 31s · 100% · 135 runs | 2h 14m · 93% · 352 runs |
| Checking by hand | 0 | 12s · 1% · 5 runs | 0 | 0 | 46s · 2% · 39 runs | 1m 03s · 3% · 13 runs | 3s · <1% · 9 runs | 10m 29s · 7% · 247 runs |
| Typecheck | 32s · 5% · 64 runs | 1m 04s · 3% · 123 runs | 44s · 51% · 85 runs | 39s · 7% · 60 runs | 1m 03s · 2% · 118 runs | 1m 26s · 3% · 121 runs | 1m 50s · 5% · 202 runs | 2m 56s · 2% · 307 runs |
| Build | 11s · 2% · 21 runs | 11s · <1% · 19 runs | 16s · 19% · 34 runs | 9s · 2% · 21 runs | 13s · <1% · 22 runs | 14s · 1% · 20 runs | 11s · 1% · 19 runs | 13s · <1% · 23 runs |
| Managing processes | 0 | 0 | 0 | 0 | <1s · <1% · 1 run | 0 | 0 | 3m 50s · 3% · 69 runs |
| Everything else | 17s · 3% · 60 runs | 4s · <1% · 62 runs | 27s · 31% · 683 runs | 5s · 1% · 47 runs | 7s · <1% · 111 runs | 6s · <1% · 73 runs | 9s · <1% · 174 runs | 21s · <1% · 609 runs |
| Total tool time | 9m 49s | 35m 05s | 1m 27s | 9m 33s | 44m 25s | 43m 05s | 34m 43s | 2h 32m |
| Tool calls | 164 | 253 | 739 | 160 | 352 | 299 | 473 | 1,264 |
Turns
How many replies each model took. Every reply is a turn: the harness sends the conversation, the model replies, usually ending in tool calls.
| Turns | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Turns | 164 | 252 | 739 | 160 | 350 | 294 | 472 | 1,224 |
Testing
Testing is where the builds differ most in time. These tables bring testing's tokens and time together from the tables above, show how each agent tests, and how often its tests found a real problem.
| Testing at a glance | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Output tokens on testing | 30,249 · 19% | 116,819 · 27% | 0 | 32,466 · 23% | 132,872 · 17% | 128,391 · 26% | 95,422 · 15% | 142,901 · 21% |
| Input tokens from testing | 27,669 · 15% | 26,621 · 5% | 694 · <1% | 37,065 · 19% | 47,287 · 5% | 16,657 · 3% | 71,109 · 5% | 76,569 · 2% |
| Time spent testing | 8m 50s · 13% | 33m 46s · 40% | 0 | 8m 40s · 9% | 43m 02s · 42% | 41m 20s · 37% | 32m 33s · 26% | 2h 25m · 31% |
| Test runs and checks | 91 | 113 | 0 | 73 | 194 | 142 | 144 | 599 |
| Test code written | 24,272 · 31 tests | 59,445 · 124 tests | 0 | 24,960 · 45 tests | 48,133 · 157 tests | 65,412 · 186 tests | 63,385 · 130 tests | 71,994 · 149 tests |
How each agent tests
Time and runs by way of testing: the agent's own suite, unit or browser, whole or narrowed to a few tests, and checking the app by hand.
| How each agent tests | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Test suite | 8m 50s · 100% · 91 runs | 33m 34s · 99% · 108 runs | 0 | 8m 40s · 100% · 73 runs | 42m 15s · 98% · 155 runs | 40m 17s · 97% · 129 runs | 32m 31s · 100% · 135 runs | 2h 14m · 93% · 352 runs |
| Unit tests, whole suite | 3m 46s · 43% · 52 runs | 22m 18s · 66% · 69 runs | 0 | 3m 18s · 38% · 40 runs | 28m 55s · 68% · 101 runs | 21m 02s · 52% · 68 runs | 18m 40s · 57% · 77 runs | 1h 10m · 53% · 209 runs |
| Unit tests, targeted | 0 | 11s · 1% · 5 runs | 0 | 0 | 21s · 1% · 14 runs | 35s · 1% · 8 runs | 0 | 1m 33s · 1% · 13 runs |
| Browser tests, whole suite | 5m 04s · 57% · 39 runs | 10m 55s · 33% · 37 runs | 0 | 4m 57s · 57% · 30 runs | 12m 32s · 30% · 36 runs | 15m 44s · 39% · 29 runs | 12m 02s · 37% · 49 runs | 39m 36s · 30% · 72 runs |
| Browser tests, targeted | 0 | 11s · 1% · 3 runs | 0 | 26s · 5% · 3 runs | 27s · 1% · 4 runs | 2m 57s · 7% · 24 runs | 1m 49s · 6% · 9 runs | 22m 35s · 17% · 74 runs |
| Checking by hand | 0 | 12s · 1% · 5 runs | 0 | 0 | 46s · 2% · 39 runs | 1m 03s · 3% · 13 runs | 3s · <1% · 9 runs | 10m 29s · 7% · 247 runs |
| The app: curl and the dev server | 0 | <1s · 3% · 1 run | 0 | 0 | 23s · 50% · 4 runs | 1m 01s · 98% · 8 runs | 0 | 6m 10s · 59% · 207 runs |
| Browser scripts and screenshots | 0 | 11s · 97% · 4 runs | 0 | 0 | 6s · 13% · 4 runs | 0 | 0 | 4m 19s · 41% · 27 runs |
| Probe scripts and the database | 0 | 0 | 0 | 0 | 17s · 37% · 31 runs | 1s · 2% · 5 runs | 3s · 100% · 9 runs | 1s · <1% · 13 runs |
| Time spent testing | 8m 50s | 33m 46s | 0 | 8m 40s | 43m 02s | 41m 20s | 32m 33s | 2h 25m |
| Tests in the suite | 31 · 6 browser | 124 · 22 browser | 0 | 45 · 9 browser | 157 · 26 browser | 186 · 51 browser | 130 · 26 browser | 149 · 34 browser |
| Seconds per unit test | 0.28 | 0.35 | — | 0.26 | 0.28 | 0.22 | 0.25 | 0.30 |
| Seconds per browser test | 3.53 | 1.42 | — | 2.21 | 1.11 | 1.06 | 0.83 | 1.28 |
| Median test run | 7s | 22s | — | 9s | 15s | 16s | 14s | 25s |
| Re-runs with nothing changed | 0 | 12 · 3m 19s | 0 | 2 · 2s | 4 · 31s | 3 · 33s | 2 · 37s | 118 · 41m 04s |
Did the tests work
What the agent's test failures turned out to be, and what its testing missed: bugs the independent acceptance suite still found at the end.
| Did the tests work | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Real bugs caught | 1 · 27s | 2 · 1m 15s | 0 | 2 · 1m 14s | 2 · 1m 36s | 5 · 1m 57s | 5 · 2m 35s | 10 · 22m 05s |
| Wrong tests | 2 · 25s | 12 · 9m 56s | 0 | 3 · 1m 59s | 20 · 10m 56s | 19 · 9m 21s | 22 · 13m 31s | 21 · 57m 10s |
| Flaky | 0 | 3 · 46s | 0 | 0 | 0 | 1 · 10s | 0 | 3 · 43s |
| Unclear | 0 | 2 · 4m 00s | 0 | 2 · 6m 00s | 4 · 8m 40s | 2 · 1m 28s | 4 · 4m 37s | 20 · 1h 48m |
| Bugs that got past | 0 | 6 | 9 | 0 | 2 | 0 | 3 | 19 |
| Critical | 0 | 0 | 3 · 33% | 0 | 0 | 0 | 1 · 33% | 8 · 42% |
| Major | 0 | 0 | 3 · 33% | 0 | 0 | 0 | 1 · 33% | 1 · 5% |
| Minor | 0 | 6 · 100% | 3 · 33% | 0 | 2 · 100% | 0 | 1 · 33% | 10 · 53% |
| Catch rate | 100% | 25% | 0% | 100% | 50% | 100% | 63% | 35% |
Cost
What each build's tokens cost at the model's list price, by part. Re-reading the cached conversation on every turn is the largest part for every build here.
| Cost | GPT-6.1 Sol | Sonnet 5.5 | GPT-6 Luna | GPT-6 Astra | Haiku 5.5 | Opus 5.5 | DeepSeek V4.1 Flash | GLM-5.3 Flash |
|---|---|---|---|---|---|---|---|---|
| Cache reads | $2.49 · 56% | $15.03 · 70% | $0.97 · 73% | $17.84 · 66% | $8.86 · 76% | $20.40 · 58% | $0.80 · 73% | $18.31 · 95% |
| Uncached input | $0.39 · 9% | $2.23 · 10% | $0.12 · 9% | $2.00 · 7% | $0.88 · 8% | $5.05 · 14% | $0.06 · 6% | $0.55 · 3% |
| Output | $1.60 · 36% | $4.32 · 20% | $0.25 · 19% | $7.19 · 27% | $1.86 · 16% | $9.85 · 28% | $0.24 · 22% | $0.35 · 2% |
| Total (list price) | $4.48 | $21.58 | $1.33 | $27.03 | $11.60 | $35.31 | $1.09 | $19.21 |
Critical and major failures
Every critical and major failure in the Scoring table, as the bug it was. Minor failures are counted there, not described.
- GPT-6 Luna (3 critical, 3 major):
- Critical, AC-19.3, AC-19.5: Blocking doesn't stop the blocker's mentions: when you mention someone you blocked, they're still notified.
- Critical, AC-26.5: When the last person's reply in a thread is deleted, nobody is told, so the deleted message stays on screen until a reload, showing that only agents' replies are left.
- Major, AC-14.1, AC-18.2, AC-19.1: The member list hides Ban, Edit roles and Block from everyone: it checks a field that a person's own profile rightly doesn't have, so none of the three can be done in the app.
- DeepSeek V4.1 Flash (1 critical, 1 major):
- Critical, AC-26.6: An agent deleting a message that mentions a person sends that person a live update: people see the effect of an agent's action.
- Major, AC-12.3: Someone else's new reply doesn't appear in an open thread (its reply count goes up, the reply never shows): the thread's slower first load overwrites the reply that arrived live.
- GLM-5.3 Flash (8 critical, 1 major):
- Critical, AC-25.2: Turning agents off for a server tells every person which agents were in it.
- Critical, AC-25.5: People see agents' replies in threads.
- Critical, AC-25.6: People are told when an agent deletes its own message.
- Critical, AC-26.1: The live reaction counts people see include agents' reactions.
- Critical, AC-26.4: A person deleting a message that only agents replied to deletes it outright, so the agents' replies lose the message they belong to.
- Critical, AC-26.6: Agents' actions send people live updates, among them unread counts and thread changes.
- Critical, AC-27.2: Taking a server off an agent's token doesn't reach connections already open: they keep getting that server's events.
- Critical, AC-27.6: An agent connection limited to some servers gets every server's missed events when it catches up.
- Major, AC-07.2: Clicking the server that's already selected empties its channel list, and no channel opens until another server is chosen.