20x plan user says Astra burns usage and won't follow instructions

rohitsabu.com @rohitsabu_Originally posted 2026.09.06

Open original on X

Is anyone else finding Astra difficult to work with? I’ve used three usage resets in two days. An overnight run exhausted 100% of my available allowance. I’m on the 20x plan. What concerns me is the combination of that consumption and recurring problems with instruction-following, coordination, and finishing work. Here are the patterns I’m seeing in Codex: 1. Constant messaging between tasks. Astra repeatedly messages other sessions, sometimes for reasons that seem random or add no useful information. Fresh sessions reach back into older ones and delegate work there. It gets confusing enough that I lose track of which session I started in, which one owns the task, and where the work is actually happening. Coordination should make ownership clearer. This makes it harder to follow my own projects. Codex also collapses thinking and communication- it’s not uncommon to see a task full of agents messaging each other and zero context in between. 2. Feedback leads to apologies and stopping. When I point out a problem, Astra can acknowledge it, apologize, and then stop working. The feedback was meant to correct its course so it could continue. Instead, I have to prompt it again to do the work already requested. Being receptive to correction needs to include acting on it. Sol was such a hard worker, a little stupid but made up for it in effort. Astra is both stupid and lazy. 3. Partial instruction-following and poor sequencing. It can feel as though most of a detailed prompt is ignored while a secondary point gets disproportionate attention. It then pursues a direction that misses the objective. Long tasks also expose sequencing mistakes: handoffs, restarts, and replacement sessions happen in the wrong order. In one case, recent tasks were archived before their handoffs were secured. The agent later acknowledged treating “idle” as “finished.” 4. Lost decisions and weak delivery. Accepted ideas and design directions don’t reliably survive into later output. I have to send the agent back through existing material to recover them. Meanwhile, audits, documentation, and passing checks can accumulate while the actual experience remains unfinished or confusing. Some design iterations have regressed from earlier work. Progress needs to mean a usable result that reflects the decisions already made. 5. Delegation and resource use. I’ve questioned complex work being assigned to smaller models that seemed poorly suited to it. An earlier activity review also flagged repeated test commands, frequent polling, and overlapping maintenance. That doesn’t establish what caused the overnight drain, but it raises questions about how much activity is useful. Users need clear usage attribution across parent tasks and workers, and agents need better judgment about repetition, delegation, and when to change approach. How about just exposing the assigned subagent’s model and effort? — These issues compound: missed instructions create rework, scattered coordination obscures ownership, and stopping after feedback requires more intervention. The result is a frustrating amount of supervision alongside heavy usage. I can’t yet separate model behavior from Codex’s orchestration or accumulated configuration. But something in the combined experience needs investigation. For a brand-new model, this has been disappointing. I’d like @OpenAI to focus on reliable instruction-following, correct sequencing, clear task ownership, continuing after feedback, preserving accepted decisions, and making usage understandable. GPT-6 Astra just doesn’t feel intelligent. I’m sure it’s great at blender and playing piano on the browser but not sure who uses it for that. Anyone else have these issue? I’m not sure if I’m just holding it wrong. I ran all the new model audits and instructions cleanups I saw recommended. Astra has lost my trust. @thsottiaux @embirico @ajambrosino @reach_vb @OpenAIDevs