Back after 5 months: what changed in how I build Tobira with AI agents
After our #1 launch in March I went quiet for almost five months, sorry for the silence.
When I came back, my old way of building with AI stopped working.
My Claude and Codex limits were gone in 2-3 days. Codex kept forgetting what the other sessions had done. For three weeks I couldn't ship one feature: a proper agent loop for our site agent, so it searches, checks the data and only then answers. I kept forwarding messages between the two AIs by hand, like a courier.
In the end I set up a harness: one shared space where every AI session, in any tool and on any computer, starts by knowing what we decided, what we tried and what's next. Within two days the stuck feature finally worked. Now my limits last the whole week.
I wanted to share this experience and discuss it here. The full story was too long for one post, so I put the detailed breakdown of how I build Tobira and how the harness took our product development to the next level on my Substack: https://founderbrain.substack.com/p/how-to-stop-hitting-your-claude-and?r=22exor
I'd love to have the discussion here, so these are the rules that mattered most:
Every decision goes into a shared log as my exact quote. Otherwise my "for example" turns into a rule a week later.
Every task gets a budget. One night Codex survived 17 compactions without losing the goal, and still burned 75% of my weekly limit solving nothing.
Hand over work before the chat is 80% full, and let the next session check what the previous one forgot.
"Done" means it works for real users. Green tests only mean "implemented".
On Friday we're relaunching here with Tobira Site Agent (if we get featured for Hypership Day).
It turns a website into an agent that talks to people and to other people's AI, and its agent loop is exactly the feature the harness finally unblocked. Thank you to everyone who supported us in March.
I'm still not a programmer, so I'd love to hear what you'd do differently and what I'm missing.
What does your AI keep forgetting, no matter how many times you tell it? And what finally fixed it?
Replies
the task budget rule is the one i’d steal. I had no cap and one of my agents saved 7 files about 2,340 times each. unfortunately, nothing stopped it, so every save worked... is your budget in tokens or time, and does the harness kill the task or do you?
@aleksandar_perak Ha, 7 files × 2,340 saves is some serious dedication 😅
Budget is a % of my weekly subscription limit, so closer to tokens than time. Each task gets its share, and a script pulls the real usage from the subscription so the agent sees how much it has already burned.
Who stops it: honestly, neither the harness nor me. The agent checks the script at checkpoints and stops itself near the cap. It either wraps up and reports what's done, or tells me what decision it needs. It can go a few points over if that finishes the task.
Which means it wouldn't have saved you: a looping agent may not reach the next checkpoint before the damage is done. For loops you need something outside the model, like a hook that counts identical tool calls and blocks after N. I only have a rule for that (two identical failures → change approach), not a hard counter yet. Did you figure out why it kept saving?