Every claim the video makes about how coding agents spend tokens, chased to a primary source. Checked 14 August 2026.
The video’s numbered examples — three candidate files, a five file budget, twenty unrelated files, thirty files for a one component bug, ten minutes of reading — are the video’s own worked instructions and illustrations rather than measurements of anything. They are listed under Not checked below.
Because the shots draw several of them as concrete meters and counts, which is how an illustration starts looking like data, the ones that could be misread as measured now carry the words “an example, not a measurement” on screen: the ten minute timeline, the thirty file field, the twenty file trace, and the narrowing search field. The five file budget does not, because it is a rule the script prescribes rather than a figure it reports.
The whole argument rests on the fact that what an agent reads is charged, not only what it writes. Anthropic’s published pricing is per million tokens and is split into input and output, with input billed on every request that carries it.
Source: Anthropic, Pricing — https://platform.claude.com/docs/en/about-claude/pricing
This is the one price that appears on screen, twice: once early, beside the file rows whose cost bars are growing, and once beside the search results being weighed against whole files. It is the Sonnet 5 input rate, $2 per million, and it is labelled as such and carries the source. Nothing else in the video states a price.
Prompt caching changes the arithmetic but not the direction: a cache write costs 1.25x the base input price on the five minute cache and 2x on the one hour cache, and a cache read costs 0.1x. Re-reading the same large file in a way that misses the cache is charged at the full input rate.
Source: Anthropic, Prompt caching — https://platform.claude.com/docs/en/build-with-claude/prompt-caching
The video says every request carries the ones before it. That is a property of the API rather than a habit of any particular tool: the Messages API holds no conversation state server side, so the client resends the whole thread each turn.
“The Messages API is stateless, which means that you always send the full conversational history to the API.”
Source: Anthropic, Using the Messages API — https://platform.claude.com/docs/en/build-with-claude/working-with-messages
The same mechanism covers the beat about generated text being fed back into the model. The
assistant’s previous turns are ordinary entries in the messages array on the next
request, so text the model produced is billed as input the next time round, and again on
every turn after that. The usage object on each response reports input_tokens and
output_tokens separately, and the input count grows with the thread.
Source: Anthropic, Using the Messages API — https://platform.claude.com/docs/en/build-with-claude/working-with-messages
The video’s closing argument, that extra context can be extra ways to be distracted rather than extra capability, is Anthropic’s own published position on building agents.
“Context, therefore, must be treated as a finite resource with diminishing marginal returns.”
“LLMs have an ‘attention budget’ that they draw on when parsing large volumes of context.”
The same piece names the degradation directly: as the number of tokens in the context window increases, the model’s ability to accurately recall information from that context decreases, a phenomenon it calls context rot, and it reports this emerging across models.
It also recommends the retrieval habit the video argues for, in the same terms: rather than loading everything up front, agents should maintain lightweight identifiers such as file paths and use them to load data into context at runtime.
Source: Anthropic, Effective context engineering for AI agents — https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Underneath that is a measured result rather than a vendor position. Liu et al. tested multi document question answering and key value retrieval and found a U shaped curve: performance is highest when the relevant information is at the very start or the very end of the input, and drops when the model has to reach for something in the middle, including on models built for long contexts.
Source: Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni, Liang, Lost in the Middle: How Language Models Use Long Contexts, Transactions of the ACL, 2024 — https://aclanthology.org/2024.tacl-1.9/ (preprint: https://arxiv.org/abs/2307.03172)
The video’s second fix is to name the directories the agent must not open. In Claude Code this is enforced by the tool layer rather than left to the model’s discretion: permission rules take deny entries for reads, and a read deny rule also blocks the edit and write tools on the paths it matches. Rules accept wildcards and project relative or absolute paths.
Source: Anthropic, Claude Code settings — https://docs.claude.com/en/docs/claude-code/settings
The video’s last fix is a short file that tells the agent where the entry points are and
what is deprecated. Claude Code reads a CLAUDE.md in the project as standing instructions
for that repository, which is the mechanism the video is describing.
Source: Anthropic, Claude Code settings — https://docs.claude.com/en/docs/claude-code/settings