Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next

Summary of technological concepts & product features (Claude Code / agentic coding)

Claude Code as an AI coding “agent”

• Claude Code is positioned as an AI coding agent that can replace much of manual coding.
• Boris’s view:
  • In many cases, coding is effectively “solved”: the model can generate 100% of code for simpler/common stacks.
  • It’s not fully solved yet for:
    • Very complex codebases
    • Uncommon / “weird” languages
  • Waiting for newer, stronger models often resolves many of these gaps.

Product “product-overhang” and early growth

• Claude Code was created inside an Anthropic incubator (“Anthropic Labs”) as an experiment.
• Early performance was weak:
  • The first ~6 months didn’t work well.
  • Growth had weak exponential characteristics even when adoption existed.
• Growth accelerated after successive model releases:
  • Opus 4 (May)
  • Then 4.5 / 4.6 / 4.7
  • These were described as inflection points that made the agent genuinely usable.

Stack choice: TypeScript + React

• The Claude Code codebase is described as TypeScript + React.
• Rationale:
  • These were considered “on distribution” for the model, improving early results.
  • As models improve, stack/framework dependency becomes less important.

Personal workflow: phone-based, many sessions/agents, scheduled automation

• Boris describes doing a lot from his phone using the Claude app.
• He runs:
  • Multiple concurrent sessions
  • Each session can spawn many agents (up to hundreds)
  • Nightly background work can include thousands of deeper tasks
• Scheduling mechanism: “/loop”
  • Uses cron-style scheduling via Claude to run repeating jobs (e.g., every minute, every 5 minutes, daily).
  • Examples:
    • Babysitting PRs: fixing CI, auto-rebasing
    • Keeping CI healthy: fixing flaky tests
    • Feedback ingestion: collecting Twitter feedback and clustering it
• Persistence via “routines”
  • A server-side equivalent that continues even if the laptop is closed.

How teams might change

• Expect more generalists, specifically cross-disciplinary generalists, such as:
  • Engineering + design
  • Engineering + data science + product
• On the Claude Code team, Boris claims everyone codes, including:
  • Engineering manager, PM, designers, data scientist, finance, user research
  • The implication: coding becomes a shared baseline skill.

“SaaS apocalypse” argument: not the usual take

• Boris predicts AI will make software writing 10–100x cheaper.
• But he argues SaaS won’t collapse as people fear.
• Two framed predictions using business “modes/powers”:

1. Some forces get weaker

   • Switching costs decrease because models can help migrate/repurpose tooling.
   • Process/workflow power becomes less differentiating as models improve at iterating process
     • He cites behaviors like “hill climbing” via instructions such as “iterate until it’s done,” referencing especially model 4.7.

2. Some forces remain important

   • Network effects
   • Scale economies
   • Cornered resources still matter.

• Startup disruption accelerates
  • He expects about 10x more disruptive startups because small teams can compete without legacy internal resistance or workflow retraining.

“Model vs product” credit and where improvements will come from

• Success is framed as a mix of:
  • Model capability
  • Product/harness details
• Early balance was around 50/50.
• As models improve, the harness becomes less critical, but it still needs evolution:
  • Make loops more first-class
  • Improve the ability to run many agents without users manually managing tool logic
• Safety burdens may reduce over time if models become reliably capable:
  • Examples mentioned: prompt injection defenses, static verification, permission modes, human-in-the-loop
  • The claim: future models may need less “heavy handling” to do the right thing.

Multi-agent parallelization and “objective / orchestration”

• On the product side, parallelization is mostly:
  • Prompting/orchestration
• Over time, models should learn delegation/parallelism more naturally.
• Loop behavior example:
  • The model detects changing data
  • Starts a scheduled loop to produce reports
  • Sends reports via Slack
  • Uses MCP integrations to do it.

Cloud vs local compute

• Boris downplays cloud vs local as a primary engineering decision.
• With agents/models improving, the model/agent will choose where to run (local vs cloud) based on task needs.
• In a couple years, models may handle:
  • code generation
  • starting agents
  • building environments making compute-location decisions less central.

General knowledge work and “computer use” / MCP

• To extend agent workflows beyond coding, he highlights:
  • MCP (Model Context Protocol) as the integration path
• Example:
  • “Co-work” can connect to business tools like Salesforce, Google Docs, Google Calendar via MCP connectors so the agent can act in real systems.
• If MCP isn’t available:
  • “Computer use” can operate through the UI of local software
  • Tradeoff: it’s slower
  • He notes Anthropic is ahead and references improvements around 4.7.

What’s coming next (teased)

• Claude Code features landing in the coming weeks.
• Continued improvement in:
  • Loop / batch / massively parallelized agents
  • Computer use

---

Reviews / guides / tutorials

• No formal “review” or step-by-step tutorial content was provided.
• However, Boris offers practical setup guidance:
  • Use /loop for recurring agent tasks (CI maintenance, PR babysitting, data collection)
  • Use routines for persistence when offline
  • Use MCP-style tool access so agents can act in real software systems

---

Main speakers / sources

• Boris Cherny (Anthropic; creator of Claude Code)
• Lauren Reader (interviewer/moderator; Anthropic)
• Mike Krieger (Anthropic; referenced as leading round-two of the team)
• Audience questions mention:
  • Dan, Jared, Jamie Nester, Ryan Sean (as named in the subtitles)
• Acquired podcast is referenced (hosts mentioned; Boris says he did an unplugged)