A 21-day agent run migrated a key legacy module to TypeScript and React for under $500 in model tokens, while exposing the visual checks needed before production.
Read in full here:
A 21-day agent run migrated a key legacy module to TypeScript and React for under $500 in model tokens, while exposing the visual checks needed before production.
Read in full here:
The interesting part here isn’t just the <$500 figure; it’s that the 21-day run surfaced the visual checks needed before production. That makes evaluation look more like a pipeline question: how much context can the agent retain between migration steps, and which observations are worth carrying forward? In my own experiments I’ve been looking at Contextpress (GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub · contextpress · PyPI), a deterministic local compressor for message lists, as a way to drop duplicate tool output while preserving decisions and constraints. I’d want to compare migrations by replaying with a fixed budget and tracking review findings, retries, and regressions—not token spend alone. The visual checks seem especially valuable because they expose failures that a clean typecheck can miss.
The <$500 headline is only meaningful once you know whether those 21 days were one long cached session or many cold starts. A cache miss re-writes history at about 2× base input instead of reading it at 0.1× — a 20× swing on the prefix — so the same migration can land closer to $200 or $800 depending on TTL gaps and how often the agent compacted. Parallelizing steps helps only if you don’t copy the growing context into each worker; otherwise N workers can cost more than one patient agent. I wrote up that parallel-agent math with current Anthropic multipliers here: Claude Code agent teams and subagents: the real token cost · AI//COST (disclosure: I work on that site).
The context question is where most of the money goes, at least in the runs we’ve measured.
We benchmark long coding tasks on Claude Code, Codex, and Oh My Pi: 32 seeded sessions per agent, same tasks with and without context reuse. Writing the code is the cheap part. The expensive part is the start of each session, where the agent re-reads files it has already seen to rebuild its picture of the codebase. On multi-session work, that re-reading was the largest single block of input tokens.
What moved the number for us was keeping a short note between sessions (file paths, the line ranges that matter, and what changed) and handing it back before the prompt, instead of letting the agent rediscover it. We call it scoped context reuse. Token use across a task fell 58.6-61.5%, depending on the agent. Cost fell less, 53.6% on Claude Code, because output tokens don’t shrink.
It also fits the cache point above. A cold start costs twice over: the cache miss itself, and the agent paying full price to re-read code it has already read once.