Field report · 3 min read
Shared memory failed its first real test
A growth trial repeated a live website change, showing that access to shared memory had not made the planner use it.
Grex’s new shared memory was supposed to stop agents from proposing work already done. Its first acceptance trial showed that the growth team still repeated a live website change.
Grex coordinates specialized AI teams, called packs. Doctrina is their curated shared memory: a record of completed work and known facts about the sites they serve.
Access to memory did not change the plan
The trial asked the growth team to use stored website facts and leave deployed changes alone. It re-proposed purchase prompts already live on HeirLogics and failed to ground a DrivewayGold draft in the supplied facts.
The failure was specific. Giving a team a memory interface had not established that its planning process consumed that memory.
Business growth missions remained suspended. A separate gap also remained: the completed-work ledger still needed manual updates when changes shipped.
Social research cleared its first checkpoint
A seven-day social mission needed three verified reply briefs in its first 24 hours. It reached five by late morning and seven by the end of the evening, against a total target of twenty-five.
The mission used seven approved feeds. Output went quiet for much of the afternoon as the available material was exhausted. The planner needed a way to vary its search without repeatedly scanning the same sources.
A reply-length fix also hit installation trouble. An unsupported package field caused repeated rollbacks. A corrected release installed under the running mission without interrupting it.
The seven briefs were prepared responses with evidence. They were not seven published replies or measured engagements.
A new discovery team started late
The discovery worker had no installed capability to read the intended startup and developer sources. A new pack, problem-radar, was built to look for problems worth investigating on Product Hunt, Y Combinator, and Hacker News.
It required source evidence before writing a brief and prohibited invented price claims. Integration exposed missing planner components and an installation dependency that worked only inside the source repository.
The team repaired those issues and launched the first three-hour mission at 11:47 p.m. Central, nearly three hours after the intended go decision. No verified result had yet established that the new capability worked.
Business signals remained preliminary
A portfolio measurement pass recorded four route-planning events for WeatherRuta and HeirLogics’ first two analytics sessions. These were usage signals, not sales.
Human engineering work also added purchase measurement to an existing checkout. That work should not be counted as autonomous Grex output.
Recorded external and mission spend for the day was $0. The strongest completed result was the social mission’s first checkpoint. The growth trial failed, and the new discovery mission still owed its first proof.