The idea in one picture
zerd moves the hard parts of an agent run into code the model cannot skip, shown problem by problem.
Most harnesses make the model do the hard parts and then hope. zerd moves the hard parts into code the model cannot skip.
| Problem every agent has | What most do | What zerd does |
|---|---|---|
| Finding the right file | grep around, read a lot | BM25 search, outline, symbol reads, a read-only scout |
| Re-reading files | send the whole file again | per-file map of lines already seen; a changed file comes back as a diff |
| Context filling up | summarize everything at 90% | newest 40k tokens verbatim, older turns folded into a summary, nothing important paraphrased |
| Model says "done" | trust it | verification gate, then an independent reviewer on a different model |
| Reviewer cries wolf | show every claim | a verifier tries to refute each finding; HIGHs are reproduced in a sandbox |
| Model invents a file or line | nothing | every path:line in the answer is checked against the tree |
| Safety | "please don't touch ~/.ssh" in the prompt | confined tools, a path-scanning shell, an OS sandbox |
| Provider quota hits | the session dies | fails over to another provider and says so |
| A bad edit | git checkout and hope | pre-run backup and /undo |
| Forgetting what you told it | start over each session | lessons learned from your corrections, written into AGENTS.md |