What changed.
And why.

They changedClaude Code v2.1.205 added an auto-mode rule that "blocks tampering with session transcript files," alongside a full /doctor configuration checkup.

So the brain nowleans on that protection: your run history stands as checkable evidence, so an agent's report of its own work arrives with receipts for free. The brain reads the live record to confirm what shipped.

Source: Claude Code v2.1.205 · shipped July 8, 2026

The evidenceClaude Code's memory docs say instruction and auto-memory files are "loaded at the start of every conversation," and warn that "if two rules contradict each other, Claude may pick one arbitrarily." Their advice: "review your CLAUDE.md files … periodically to remove outdated or conflicting instructions."

So the brain nowruns a standing consolidation pass over its own notes: merge duplicates, retire dead rules, reconcile contradictions, date-stamp what stays. Anything loaded every session stays current by editing, not by piling up.

Source: Claude Code memory docs · living reference, consulted July 2026

The evidenceSimon Willison's rule for coding agents is blunt: "Never assume that code generated by an LLM works until that code has been executed," and if it was never run, trusting it is "pure luck." An agent reporting that a task is done is not proof the task is done.

So the brain nowtreats an agent's "done" as a claim, not a fact. Separate, independent verification runs before work counts as finished: what a builder reports and what a fresh check confirms are two different things.

Sources: Simon Willison · Agentic manual testing · March 6, 2026 · First run the tests · February 28, 2026

They changedClaude Code v2.1.198 made subagents "run in the background by default, so Claude keeps working while they run." Parallel agent work went from an opt-in to the default behavior.

So the brain nowtreats collision protection as the baseline, not the edge case: every parallel task gets its own isolated copy of the project, so two agents can never write to the same file. If agents run in parallel by default, so should their isolation.

Source: Claude Code changelog v2.1.198 · shipped July 1, 2026

They changedan export-control directive suspended API access to two frontier Claude models, Fable 5 and Mythos 5, in mid-June, days after their launch. Access was restored on July 1, about three weeks later.

So the brain nownever assumes a top model is permanently available. Model routing carries a fallback chain: when the first choice disappears, work degrades to the next-best model instead of stopping, and access gets re-checked rather than presumed.

Sources: Simon Willison · June 13, 2026 · anthropic.com/news · July 1, 2026

They changedClaude Sonnet 5 launched on June 30 as the new default model (shipped in Claude Code v2.1.197). Its tokenizer produces roughly 30 percent more tokens for the same text, and its introductory pricing ($2/$10 per million tokens) ends August 31, when standard pricing ($3/$15) takes over.

So the brain nowrecalculates how much conversation fits in one working session under the new tokenizer, and logs August 31 as a fixed date to revisit cost decisions before the price changes. Any session budget tuned before June 30 undercounts by about a third.

Sources: anthropic.com/news · June 30, 2026 · platform release notes · June 30, 2026

The evidenceClaude Code's docs make model choice the real cost lever: "Sonnet handles most coding tasks well and costs less than Opus. Reserve Opus for complex architectural decisions or multi-step reasoning." The subagent docs add you can "control costs by routing tasks to faster, cheaper models like Haiku," and Claude Code has let subagents "dynamically choose the model" since v2.0.28 (October 2025).

So the brain nowtreats which model does the work as the lever that matters: routine work goes to the cheaper model by default, and only a named risk (a failure mode no automated check would catch, real design judgment, or wide blast radius) escalates to the stronger one.

Sources: Claude Code · Manage costs · Subagents · living reference, consulted July 2026

They changedClaude Code v2.1.183 made auto mode block destructive commands it was not asked for: "git reset --hard, git checkout -- ., git clean -fd, git stash drop … blocked when you didn't ask to discard local work," and "terraform destroy / pulumi destroy / cdk destroy … blocked unless you asked for the specific stack."

So the brain nowdraws the same line by default. An agent gets room to build, but the irreversible moves are gated behind an explicit ask: throwing away local work, tearing down infrastructure. Autonomy on the safe operations, a hard stop on the ones you cannot undo.

Source: Claude Code v2.1.183 · shipped June 19, 2026

They changedOver ten weeks, Anthropic clarified how automated tooling should authenticate: a block on third-party tools using a personal subscription login (April 4), a plan to route agent workloads onto separate metered credits (announced May 13), then a pause of that plan on the day it was to take effect (June 15).

So the brain nowbuilds on the direct, metered connection, the path Anthropic points automated work toward. Wiring there by default keeps your automation on supported, first-class ground from the start.

Sources: Hacker News · April 4, 2026 · The New Stack · May 13, 2026 · The New Stack · June 15, 2026

The evidenceClaude Code ships worktree isolation for exactly this problem. Its docs: "Running each Claude Code session in its own worktree means edits in one session never touch files in another," and the --worktree flag plus subagent isolation: "worktree" shipped in v2.1.49 (February 2026).

So the brain nowgives every parallel agent its own isolated checkout, so two agents can never write over each other's files. When agents run at the same time, isolation is the default, not the exception.

Sources: Claude Code · Worktrees · living reference, consulted July 2026 · v2.1.49 · shipped February 19, 2026

The evidenceClaude Code's cost docs state the mechanism plainly: "Token costs scale with context size: the more context Claude processes, the more tokens you use." And the fix: "Clear between tasks: Use /clear to start fresh when switching to unrelated work. Stale context wastes tokens on every subsequent message."

So the brain nowruns short, focused sessions per task and clears between them, keeping continuity in durable notes outside the conversation. A long thread is the expensive way to remember.

Source: Claude Code · Manage costs · living reference, consulted July 2026

The evidenceClaude Code's docs are explicit that a written rule is not enforcement: "Claude treats them as context, not enforced configuration. To block an action regardless of what Claude decides, use a PreToolUse hook instead," because hooks "apply regardless of what Claude decides to do." Hooks shipped in Claude Code v1.0.38 (June 2025).

So the brain nowwires its riskiest rules into hooks that fire at the moment a rule would be broken: a messy starting state, two tasks on one file, a note missing its required tag. A rule that lives only in a memory file is a suggestion; a hook is a wall.

Sources: Claude Code · Memory · Hooks · living reference, consulted July 2026 · v1.0.38 · shipped June 30, 2025

The evidenceClaude Code's docs confirm the cost of loading everything up front: "CLAUDE.md files are loaded into the context window at the start of every session, consuming tokens," and those tokens "are present even when you're doing unrelated work." The guidance is to keep the file "under 200 lines" and move specialized instructions out to on-demand skills. Claude Code has supported CLAUDE.md imports since v0.2.107 and path-scoped rules since v2.0.64, both for exactly this.

So the brain nowkeeps live instructions lean and separate from settled history, so an auto-loaded file only carries what every session actually needs. Every line loaded at startup is paid for on every startup.

Sources: Claude Code · Memory · Manage costs · living reference, consulted July 2026

They changedClaude Code v2.1.154 introduced dynamic workflows: "ask Claude to create a workflow and it orchestrates work across tens to hundreds of agents in the background." The same release shipped Opus 4.8, whose fast mode runs "at a fraction of its previous cost: 2x the standard rate for 2.5x the speed."

So the brain nowplans for fan-out as normal, not exotic: many small jobs at once, each in its own isolated copy so they cannot corrupt each other, with model and speed tier chosen per job. A latency-sensitive loop can pay a little more for a fast tier; a bulk pass takes the cheap one.

Sources: Claude Code v2.1.154 · shipped May 28, 2026 · anthropic.com/news · May 28, 2026

The evidenceSimon Willison marked the point where coding agents hit product-market fit, and noted that organizations were caught off guard by how much running them costs at scale.

So the brain nowkeeps the meter in view. It surfaces the running cost of a piece of work up front rather than after the invoice, because a tool that quietly spends is a tool you stop trusting.

Source: Simon Willison · May 27, 2026

The mechanismwhen two tasks write the same file at the same time, the writes do not merge. Whichever saves last wins, and the other's unfinished work is silently overwritten, with no conflict raised to warn anyone.

So the brain nowchecks every new task's file list against everything already in flight and refuses to run two tasks on the same file. A collision is cheaper to block before it exists than to untangle once it is in the history.

Self-evident mechanism. No external source claimed.

The mechanisman agent editing a file and a live server serving that file are two separate facts. Between the edit and what a browser finally receives sit build steps, caches, and proxies, any of which can keep serving the old bytes after a deploy reports success.

So the brain nowconfirms the actually served file contains the actual change before anything counts as done. Verification reads the live bytes, not the deploy log.

Self-evident mechanism. No external source claimed.

The evidenceSimon Willison flagged that Claude Code briefly showed up as Max-plan-only on the pricing page, moved off the $20 Pro tier, before the change was reversed within hours.

So the brain nowtreats pricing and plan tiers as live state, not settled fact. It re-checks the current pricing page before it reasons about cost, instead of trusting what was true last quarter.

Source: Simon Willison · April 22, 2026

They changedAnthropic shipped Claude Opus 4.7 with a new tokenizer that produces roughly 1.0 to 1.35 times as many tokens for the same text.

So the brain nowre-baselines its token estimates whenever the tokenizer changes. A budget or a context plan tuned to the old tokenizer silently undercounts, so the count gets recalibrated, not assumed. The same discipline caught the Sonnet 5 change two months later.

Source: anthropic.com/news · April 16, 2026

The evidenceOn Claude Code's auto mode, Simon Willison drew a sharp line: an AI classifier deciding whether a command is safe is not the same as a deterministic sandbox that cannot run the command at all.

So the brain nowuses a model-based safety check as a second layer, never the only one. A hard constraint (an isolated copy, a permission wall, a blocked command) does the real containment; the model's judgment is a backstop on top, not a substitute.

Source: Simon Willison · March 24, 2026

They changedClaude Code v2.1.59 launched auto-memory: "Claude automatically saves useful context to auto-memory. Manage with /memory."

So the brain nowkeeps a durable memory too, but a deliberately curated one. Auto-memory is an in-tool log the agent writes as it goes; the brain's memory is maintained on purpose, pruned and reconciled, so a stale note gets retired instead of accumulating. Different mechanism, different guarantee.

Source: Claude Code v2.1.59 · shipped February 25, 2026

They changedAnthropic's terms, reported by The Register on February 20, made explicit that a personal subscription's OAuth tokens may not be used inside third-party products, including through the Agent SDK. Anthropic had begun enforcing that pattern earlier in 2026.

So the brain nowruns automated work on a metered API key, the authentication Anthropic supports for third-party tools. Bringing your own key keeps you on supported ground, and it is the stance the brain adopted early and has held since, roughly four months before the metered-credits changes of mid-June.

Source: Claude Code legal and compliance docs · reported by The Register, February 20, 2026

They changedClaude Code v2.1.32 shipped Opus 4.6 and a research-preview "agent teams feature for multi-agent collaboration," where several agents coordinate on one task.

So the brain nowruns multiple agents on one job as a normal shape, with an honest difference worth stating: agent teams share a single working tree; the brain isolates at the file level instead, giving each agent its own copy. Both coordinate work; the brain's guarantee against two agents clobbering one file is stronger.

Sources: Claude Code v2.1.32 · shipped February 5, 2026 · anthropic.com/news · February 5, 2026

They changedClaude Code v2.0.60 "added background agent support. Agents run in the background while you work."

So the brain nowtreats hand off and keep going as the normal way to build: work is dispatched to a background agent, the front desk stays free to think and plan, and finished work comes back to be checked. The person in front never has to sit and watch a build.

Source: Claude Code v2.0.60 · shipped December 5, 2025

They changedAnthropic released Claude Opus 4.5 at $5/$25 per million tokens, a roughly threefold price cut on the top tier (shipped in Claude Code v2.0.51).

So the brain nowre-examines routing rules whenever the price of a tier moves, not just when a new model ships. A rule that avoided the flagship purely on cost can become wrong overnight when that flagship gets three times cheaper. Cost assumptions expire.

Sources: anthropic.com/news · November 24, 2025 · v2.0.51 · shipped November 24, 2025

They changedAnthropic introduced Claude Skills, packaged expertise the agent loads only when it is relevant (added to Claude Code in v2.0.20).

So the brain nowpackages specialized know-how as skills that load on demand instead of riding in context all the time. A workflow the brain uses once a week does not tax every unrelated session; it arrives when it is needed and costs nothing when it is not.

Sources: claude.com/blog · October 16, 2025 · v2.0.20 · shipped October 16, 2025

They changedClaude Code v2.0.12 released the plugin system: "extend Claude Code with custom commands, agents, hooks, and MCP servers from marketplaces."

So the brain nowis built to be installed and carried, not hand-wired per machine. Its commands, workers, enforcement rules, and connections travel as one bundle, so a fresh setup gets the whole working style at once instead of rebuilding it piece by piece.

Sources: Claude Code v2.0.12 · shipped October 9, 2025 · anthropic.com/news

They changedClaude Code 2.0 landed with a native VS Code extension, "/rewind a conversation to undo code changes," and a "/usage command to see plan limits," alongside Anthropic's post on making Claude Code work more autonomously.

So the brain nowassumes an editor-native home and a local undo, with one honest caveat it keeps front of mind: an in-tool rewind undoes edits, not a deploy or a pushed commit. Local undo is a convenience; it is not a substitute for verifying what actually shipped.

Sources: Claude Code v2.0.0 · shipped September 29, 2025 · anthropic.com/news · September 29, 2025

They changedClaude Sonnet 4 gained a 1 million token context window, with a premium on requests over 200K tokens ($6/$22.50). In March 2026 that long-context premium was removed.

So the brain nowtreats a bigger context window as a tool, not a habit. A million tokens is available when a task genuinely needs the whole picture, but the default stays lean; and since the premium that once punished long context is gone, the brain no longer contorts work just to dodge a surcharge that no longer exists.

Source: claude.com/blog · 1M context · August 12, 2025 · premium removed March 2026

They changedClaude Code v1.0.60 added custom subagents: "you can now create custom subagents for specialized tasks."

So the brain nowruns specialized workers with narrow jobs: a read-only scout that investigates without touching anything, a builder that writes code, a reviewer that checks it. Each keeps its own context, so a heavy search never floods the conversation that has to make the decision.

Source: Claude Code v1.0.60 · shipped July 24, 2025

The evidenceChroma's "Context Rot" research showed that model performance degrades non-uniformly as input length grows: a longer context is not just more expensive, it can quietly get less accurate.

So the brain nowtreats a long context as a quality risk, not only a cost. It clears at task boundaries and keeps working context tight, because the failure mode is not just a bigger bill; it is a subtly worse answer that nothing flags.

Source: Chroma · Context Rot research · July 14, 2025

They changedClaude Code reached general availability alongside Claude 4: "Claude Code is now generally available," "introducing Sonnet 4 and Opus 4 models" (Opus 4 at $15/$75, Sonnet 4 at $3/$15).

So the brain nowhas a stable home platform to build on, and a founding routing rule: Sonnet 4 as the everyday workhorse, Opus 4 reserved for long-horizon, high-stakes work where its cost earns out. The default is the affordable tier; the expensive one is a deliberate choice.

Sources: anthropic.com/news · May 22, 2025 · Claude Code v1.0.0 · shipped May 22, 2025

They changedClaude Code v0.2.93 made sessions resumable ("resume conversations from where you left off with claude --continue and claude --resume") and gave Claude a Todo list "that helps it stay on track."

So the brain nowtreats the terminal as disposable and the work as durable. Continuity lives in saved state and written notes, not in one fragile open window, so closing a session loses nothing and a cold start is as capable as a warm one.

Source: Claude Code v0.2.93 · shipped April 30, 2025

They changedClaude Code v0.2.47 added "automatic conversation compaction for infinite conversation length," which summarizes older history to keep a session going.

So the brain nowtreats compaction as lossy and avoids leaning on it. Summarizing a long thread throws away detail every time it runs, so the brain prefers to close a task and start the next one fresh from durable notes, rather than let one endless conversation slowly blur.

Source: Claude Code v0.2.47 · shipped March 18, 2025

They changedClaude Code v0.2.44 exposed reasoning depth as a plain instruction: "say 'think' or 'think harder' or even 'ultrathink'."

So the brain nowspends deeper reasoning deliberately, not by reflex. Thinking budget is a real cost, so it is reserved for genuine forks and hard design calls, and kept off routine, mechanical work where extra deliberation buys nothing.

Source: Claude Code v0.2.44 · shipped March 15, 2025

They changedClaude Code v0.2.31 added custom slash commands: "markdown files in .claude/commands/ directories now appear as custom slash commands to insert prompts into your conversation."

So the brain nowcaptures a repeated instruction once, as a named command, instead of retyping it. A workflow you run often becomes a single word, which keeps the phrasing consistent and turns hard-won prompt craft into reusable tooling.

Source: Claude Code v0.2.31 · shipped March 5, 2025

They changedAnthropic launched Claude 3.7 Sonnet and, with it, the first research preview of Claude Code. Extended thinking arrived, with thinking tokens billed as output.

So the brain nowtreats reasoning as something you pay for by the token and budgets for it. Extended thinking is worth its cost on the hard problems and wasteful on the easy ones, so the brain matches the depth of thought to the difficulty of the task rather than leaving it on full everywhere.

Source: anthropic.com/news · February 24, 2025

They changedAnthropic launched the Model Context Protocol, an open standard for connecting models to external tools and data.

So the brain nowreaches other systems through a standard tool interface, not one-off glue. A data source or an external service is wired in as a first-class tool the agent can call, which keeps integrations uniform and keeps their heavy definitions out of context until they are actually used.

Source: anthropic.com/news · November 25, 2024

They changedAnthropic shipped an upgraded Claude 3.5 Sonnet and Claude 3.5 Haiku at $0.80/$4 per million tokens, plus the first computer-use capability.

So the brain nowkeeps a genuinely cheap, fast tier in the rotation for high-volume work. Bulk, mechanical, or repetitive subtasks route down to the small model, which frees the expensive tiers for the judgment calls that actually need them.

Source: anthropic.com/news · October 22, 2024

They changedAnthropic released the Message Batches API: submit work asynchronously and get it back within 24 hours at 50 percent off.

So the brain nowseparates interactive work from bulk work, and routes anything that does not need an immediate answer to the cheaper asynchronous path. Half price for work that can wait is a default worth taking, not an optimization to remember later.

Source: anthropic.com/news · October 8, 2024

They changedAnthropic added prompt caching to the Claude API, cutting cost on repeated context by as much as 90 percent and reducing latency.

So the brain noworders every prompt so the stable parts come first: fixed instructions and shared context ahead of the part that changes. That layout lets the cache do the heavy lifting on everything the agent re-sends, which is most of a long session.

Source: claude.com/blog · Prompt caching · August 14, 2024

They changedAnthropic released Claude 3.5 Sonnet at $3/$15 per million tokens, a mid-tier model that outperformed the previous flagship, Claude 3 Opus, at a fraction of the price.

So the brain nowholds a rule that starts right here: default to the capable workhorse tier, and reserve the top tier for the work that truly needs it. The moment a mid-priced model beat the old flagship, "always reach for the biggest model" stopped being the smart default. It never came back.

Source: anthropic.com/news · June 20, 2024