← sources

Opus 5.5 Ultracode is like hiring all four of the big four, in the worse way.

source: reddit (reddit:AI_Agents)published 2026-09-29 08:19ingested 2026-09-30 02:37

retained text

Sorry for the rant reddit, I just need to get this off my chest.

I've been building with Opus 5.5 on Ultracode for the last week.  It's has taken ages to get my features built and deployed.

Every spec the f$%ker wants to argue over and bring in 2-3 sub agents to do multiple reviews at EVERY F$%KING turn.

Then it add's the multitutudes of minute guard rails and checks, that no way would you do in a real life human built deveopment project.  It's like there's an army of graduate developers who all want to get their say on the every point.

For each requirement and spec, rather than iterating through the development, testing and veifying the requirements it goes off down a rabbit hole and introuduces massive scope creep.

It's like when you get one of the big consultancies on site and the embedded consultant identifies "adjacent problems" or missing capabilities that you have to hire more of their specialists for.

Every time Opus 5.5 does some research, builds or deploy's it brings in another 2-3 sub agents that run for 10 to 20 minutes.

This is from my last session, which has taken me all day.  I use spec driven development, so everthing has already been spec'd... I asked Claud to provide the data.

\-------

Here are all the subagents run since we started testing this morning. Times are local (AEST). "Ran for" is minutes:seconds.

|Workflow|Agent|Started|Ran for|Tokens|Tool calls|
|:-|:-|:-|:-|:-|:-|
|**Spec row 187** (embedder null vectors)|draft|09:49|17:13|287,514|62|
||verify-code-truth|10:06|11:33|207,227|56|
||verify-buildability|10:06|11:01|193,425|36|
||revise|10:18|7:06|173,806|53|
||backlog|10:25|1:44|109,318|13|
|**Row 149 correction + spec row 188**|correct-149|11:00|2:43|107,571|17|
||draft-188|11:00|15:44|282,859|60|
||verify-truth-188|11:16|11:12|183,469|67|
||verify-build-188|11:16|9:52|175,172|58|
||revise-188|11:27|5:21|125,792|53|
||backlog-188|11:32|2:17|93,769|19|
|**Spec row 189 + limits sweep**|sweep-limits|13:38|6:20|161,223|55|
||draft-189|13:45|35:28|482,641|96|
||verify-truth-189|14:20|19:08|293,319|81|
||verify-build-189|14:20|14:51|222,842|32|
||revise-189|14:39|15:49|272,222|91|
||record-189|14:55|4:09|166,574|27|
|**Fact-check the features guide**|check-1-2.2|16:07|9:40|185,956|68|
||check-2.3-2.4|16:07|7:40|144,891|69|
||check-2.5-2.8|16:07|6:47|154,071|63|
||check-3-5|16:07|9:43|202,769|75|
||apply|16:17|9:45|269,033|60|
|**Promotion briefing** (153/165/161/163)|brief-153|17:51|3:27|155,127|22|
||brief-165|17:51|3:17|142,636|16|
||brief-161|17:51|3:33|147,417|21|
||brief-163|17:51|3:53|150,289|21|
|**Record the demo decisions**|record|17:57|5:47|182,658|32|
||check|18:03|2:22|107,533|17|
|**Build row 153**|build:153|18:06|*running (3:30 so far)*|—|—|

* **Totals (finished):** 28 agents, about **4 h 17 min of agent time** and **5.4 M tokens**. Agents that started at the same time ran in parallel, so the wall-clock time was shorter.
* **Wall-clock per workflow:**

|Workflow|Wall-clock|
|:-|:-|
|Spec row 187|\~38 min|
|Row 149 correction + spec 188|\~35 min|
|Spec row 189|\~81 min, the longest: its draft alone took 35 min|
|Fact-check the guide|\~20 min|
|Promotion briefing|\~4 min|
|Record the decisions|\~8 min|

* **Not in the table:** during the row-by-row walkthrough itself, I ran the checks directly with no subagents. That includes the real-browser comparison of old and new code for row 149. The overnight run's agents are also excluded.

\-----

Maybe it's because I have too many Matt Pocock skills installed....



The days I'm essentially a technical presales and product manager.  I speak to the custmers so I know what to get the agents to build. I do not know if I have the patience to use models like this.

Take me back to Opus 4.6 and GPT 5.4

document fetched 2026-09-30 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wt4go2/opus_55_ultracode_is_like_hiring_all_four_of_the