An agent edit you can't classify is a guess you're merging. The Engineering tier is the safety net: every change checked and receipted — contract, effects, behavior — across 138 grammars, in any MCP agent. And it answers in one call instead of a dozen file reads.
unsafe_surface, effect_summary, data_flow across 138 grammars.
The verify_* suite — every refactor your agent runs comes with a receipt that proves contract, effects, and behavior survived.
Claude Code, Cursor, Cline, Codex, your own Agent SDK build.
What verify_contract_preserved returns for a single change — the hunk classified, each contract dimension checked, the effect delta named.
hunk Behavior (not Rename / Move / Signature / Format) contract Preserved ├─ control preserved same branch / loop / return shape ├─ raises preserved {ValidationError} unchanged ├─ guards changed added: qty > 0 (new precondition) └─ effects preserved reads:[cart] writes:[] — no new I/O verdict review 1 dimension changed — surfaced, not blocked
A representative verify_contract_preserved receipt.
Ask your agent "is this function safe to delete?" It greps, reads 12 files looking for callers, opens 4 more to check effects. ~80K input tokens, per task.
~80K tokensOne effect_closure MCP call returns the transitive effect
signature, frontier, and call sites. ~3K tokens. 27× less.
A clean context window isn't just cheaper — an agent that isn't drowning in file reads reasons better and writes better code.
Per-task figures above are measured from our agent-task benchmark. For aggregate savings across 10 operations, see the 166-sample savings benchmark — median 85% fewer tokens.
One native binary, every grammar, zero setup. Builder Edition is free.
Download — install in 60sShip notes and new capabilities, occasionally. No spam; unsubscribe anytime.
Architectural concerns? Tools for tech leadership → /for/architects.