pranavmishra
FDE @ Harper(YC 25)
- Joined
- Nov 18, 2026
Comments
[Weekly] What are you building this week?
coolio, interesting stuff
[Weekly] What are you building this week?
the minimal task store idea really resonates. something i've run into with coding agents is that they can preserve a perfectly coherent task summary while losing track of what's actually been verified. i've been experimenting with carrying file hashes and observed check results across compactions rather than trusting completion claims.
does tracker distinguish agent-reported progress from independently verified state, or is that intentionally left to the agents?
[Weekly] What are you building this week?
this is a clever way to solve the cold-start problem. one subtle issue though: the model's initial ranking also determines which comments get enough exposure to collect votes, so the eventual human signal isn't entirely independent of the judge.
have you considered reserving a bit of randomized exposure to measure whether the model's quality judgments actually correlate with reader preferences?
[Weekly] What are you building this week?
the common harness across that many surfaces is a really interesting problem. i've been building a local coding workspace on OpenCode, and one thing i've learned is that session continuity and verified task state are two very different things. a session can resume perfectly while its previous test results are already stale.
are you centralizing execution state and verification in the harness, or letting each surface manage some of that independently?
[Weekly] What are you building this week?
been building KRYN, an open-source coding workspace running Qwen3.5-9B locally on Apple Silicon via oMLX + OpenCode.
the interesting challenge isn't getting a 9B model to write code. it's making it reliable across long-running tasks: preserving file/check evidence through compaction and restarts, managing memory pressure, and independently verifying outcomes instead of trusting the agent's "done" message.
one surprising result: stabilizing the system prompt cut new input tokens by 60%, but strict task acceptance dropped from 3/4 to 2/4 in a small dev eval. reverted it.
curious how others building agent harnesses are measuring actual reliability beyond benchmark scores and successful tool calls.