hebel. today · 9 November 202620 / 74
Agentic Engineering · AI engineering
Monaco builds custom workspaces to keep agent experiments reproducible
Mihail Eric at Monaco maintains a rule of thumb: buy tools to learn, but build them to keep. While the team evaluated commercial code review products, they built their own agent workspaces because commercial…
From 5 newsletters · 1 min read · one thing to try
Mihail Eric at Monaco maintains a rule of thumb: buy tools to learn, but build them to keep. While the team evaluated commercial code review products, they built their own agent workspaces because commercial offerings lacked the specific cloud configurations needed to reproduce failures and test patches (Latent Space). This infrastructure allows their agent, Hal, to investigate broken production code and open a pull request, while simultaneously providing salespeople with an interface to query product functionality (Latent Space).
Teams now enforce deterministic verification loops to prevent agents from relying on their own performance accounts, a shift from the ad-hoc configurations common in early agent adoption (TLDR DEV, Akos Komuves from Bitsy). Miguel Otero Pedrido implemented this at The Neural Maze by embedding hard stopping conditions—such as maximum iteration counts and USD budgets—directly into Inngest-based durable execution steps to stop agents from drowning in redundant search results (The Neural Maze). These constraints ensure that an agent stops immediately when a research round yields no new information, preventing the token waste seen in unmonitored loops.
Engineers at OpenAI scaled their Git infrastructure by building Pando, a backend that handles pushes to unrelated branches independently and reuses cached packfiles to reduce median push latency by nearly 80%. This architecture allows for horizontal scaling of writes, though it accepts weaker consistency across branches as a tradeoff for saturating client network connections.
Try today · 30 min
Implement a hard stop for your agent's research loop using a maximum iteration count and a token budget.
Define a counter variable that increments each loop and a budget check that triggers an exit when exceeded.
You need
- Current agent loop code
- Counter variable
- Budget limit constant
Steps
- Initialize an iteration counter at zero.
- Add a check at the start of your loop to compare the counter against your maximum.
- Add a conditional check to compare current token usage against your budget.
- Raise an exception or return a result if either limit is hit.
- Increment the counter at the end of each successful loop iteration.