State of Agentic Coding #7 with Armin and Ben

Technological concepts & agentic coding realities discussed

Agentic coding “loops” vs terminal workflows

• Ben and Armin discuss how better model quality has made agentic coding substantially more capable over the past year.
• Despite this, they’re skeptical that most developers are using true always-on / fully autonomous “looping” systems.
• They reference the idea from Anthropic/OpenAI-adjacent practitioners that the goal shouldn’t be “prompting more,” but architecting systems that run loops and converge over time.
• Even so, they claim many practical workflows still resemble:
  • human-in-the-loop terminal usage
  • not Minority Report-style UI
  • not large-scale autonomous agents

Quality, QA, and evaluation problems remain

• Even with bigger context windows, LLM code generation can still silently regress features or UI behavior.
• Comprehension/verification doesn’t automatically scale with code generation output.
• It’s difficult to QA what agents changed and why.
• Armin suggests that in “looping” environments, the bottleneck may shift to:
  • human comprehension, not model ability

“Psychosis” / warped reasoning

• They describe “AI psychosis” less as dramatic hallucination and more as subtle perspective warping:
  • agents generate plausible but not comprehensive hypotheses
• Example: agents propose many checks but miss the real root cause

Security research + token economics

• They connect AI coding’s rise to more automated security reporting and weaponization:
  • Security tooling finds issues at scale, but teams face a “base budget” of security triage/research.
• They claim a feedback loop risk may be emerging:
  • some open-source maintainers are getting repos “fixed” by bots/agents
  • based on issues submitted by other AI tools
  • creating changes and trust risks

---

Product/platform changes (Claude/Copilot/agent tools) emphasized

Token subsidies are unwinding

• Copilot is shifting from subscription-like usage to usage-based pricing, with bundled usage reportedly reduced.
• Claude-P / CLI agent execution is highlighted:
  • Claude-P is no longer covered by the prior subscription model.
  • Instead, Anthropic provides separate “programmatic agent credits” for CLI/SDK-style usage (including Claude via CLI/SDK, and potentially “ultra mode/workflow”).
• They interpret this as economic steering:
  • features that drive higher spend are promoted
  • they speculate providers value human-in-the-loop traces more than raw machine-to-machine execution, because those traces provide better training signal and product differentiation

Token spend becomes a dominant business metric

• The industry increasingly optimizes for token usage, since token spend correlates strongly with revenue.
• They compare this to past “metrics” (MAU/DAU), but argue tokens are more directly monetizable.

---

Reviews / guides / practical experiments mentioned

Local open-weight models progress

• They discuss experimentation progress for running models locally:
  • hardware + quantization improvements sometimes allow local models to support coding-agent-style workloads at performance approaching cloud agents from ~a year earlier
• Example: Dwarf Star 4 (DS4) (DeepSeek-derived)
  • described as an end-to-end packaged local coding model experience
  • uses quantization and on-disk caches optimized for agentic coding
  • reported performance claims:
    • prefill: ~450 tokens/sec
    • generation: ~25–26 tokens/sec
    • performance reportedly stays relatively stable until a large fraction of context is filled
• Practical constraints:
  • users may need high-end machines
  • battery/heat can become limiting
• “Flash” vs “Pro” variants:
  • flash is easier to run but requires more supervision

Claude/Coding agent behavior still needs supervision

• They recount attempts where agents refactor aggressively and produce unwanted changes, requiring careful review and rollback.

---

“Slop fork” / library rewrite discussion (Bun → Rust; performance + strategy)

“Slop forks” and token justification

• The talk revisits the idea that some projects are being “slop-forked”:
  • they cite Bun
  • a claim that the Bun creator’s company was acquired by Anthropic
  • Bun being rewritten via agentic tools (“Robun”) from Zig to Rust
• They note the rewrite isn’t fully released yet, but prebuilt artifacts exist.

Why rewrite at all if it’s already successful?

• They acknowledge uncertainty (performative vs technical reasons), but suggest:
  • evidence models can port large codebases while tests protect behavior
  • a way to prove feasibility for enterprises doing legacy migrations

Contrast with typical adoption

• Some forks (e.g., “V-next”-style for Next.js) haven’t gained huge adoption, implying wholesale rewrites differ from incremental forks.

---

Language debate: Ruby/Zig “human languages” vs AI efficiency

• They argue LLMs tend to select languages that are easier for machines, even if they introduce worse trade-offs for humans.
• Ruby is called “worst for AI” due to:
  • non-local state
  • heavy runtime metaprogramming
• Zig is framed as “human-friendly,” but hard to guarantee stability in fully machine-driven contexts:
  • memory management
  • crash tendency
• Core takeaway: they believe the ecosystem is gradually eliminating languages designed primarily for humans in favor of machine-optimized choices.

---

Key analysis claims (what’s working / what’s not)

Software volume doesn’t mean user value

• They emphasize that many apps look like “agent harness multiplexers” built on strong UI/primitive libraries (diffs/trees/ghosty-like terminal views), not necessarily agentic innovation.
• Their argument: primitives matter more than agent wrappers.

Token factory vs sustainable open-source primitives

• They claim most money goes to token sellers, while primitive tool makers often monetize indirectly or through other products.

Fast fashion / shallow conviction

• Software built with AI may get discarded quickly if it doesn’t gain early traction.
• They encourage longer project commitment.

---

Named speakers/sources (as stated in the subtitles)

• Ben Vingar (Ben Vinegar)
• Armin (Armen Aroner / “Armaroner” in subtitles)

Additional contextual mentions:

• Boris Churnney (Anthropic)
• Peter Steinberger (OpenAI/OpenCL; “OpenClaw” referenced)
• and others including Jenssen, Copilot/Anthropic, Modem/Diffs, and Dwarf Star authors