Where this is going
A short closing chapter: current frontier, honest limits, what to watch.
What models can reliably do today vs. a year ago
In mid-2025, Claude could draft code for a task you described. By mid-2026, it can navigate an unfamiliar repository, propose an architectural change, write it, test it, and commit it, often without you intervening. A year ago, “multi-turn reasoning over a codebase” was aspirational. Today it’s the default mode of working.
The pace isn’t slowing. A reasonable reading of the trend: tasks that feel like they require an engineer in 2026 will feel like they require a product manager in 2027. The constraining factor is increasingly the clarity of the person driving, not the capability of the model.
What harnesses can do today vs. a year ago
The harness has been the bigger unlock than the model, in some ways. A year ago, Claude Code was a capable REPL. Today it’s a platform: Skills compose, sub-agents parallelize, MCP plugs in new capabilities per session, agent teams coordinate. The harness has become the surface where most of the interesting engineering is happening, rather than the model itself.
Anthropic’s own writing backs this up: on the same tasks, the difference between Claude Code’s harness and other harnesses is tens of percentage points of capability. The harness is not a thin UI layer.
The honest limits
Where this stuff still falls down, routinely:
- Very long horizons. A task that takes many hours of focused work, where the context keeps growing, still tends to degrade. Compaction helps but doesn’t fully solve it.
- Novel problems with no analogs in training. The model is excellent at recombining patterns it has seen; genuinely new problem shapes stretch it thin.
- Strict determinism. Same prompt, same context, different run → slightly different output. For anything that needs bit-exact reproducibility, you still wrap Claude’s output in deterministic logic.
- Physical-world reasoning. Robotics, embedded systems, anywhere the feedback loop is slow and unforgiving: the fast iteration loop that makes LLM development work is absent.
None of these will hold forever. All of them hold today.
Near-term changes worth tracking
- Longer effective context: not just token counts, but attention quality across larger windows.
- Better self-improvement loops: from reflection-based methods toward self-play and generated training data.
- Harness consolidation: the current fragmentation (Claude Code, Cursor, Codex, Copilot, etc.) will probably converge on a few dominant designs.
- IDE evolution: see below.
- Agent teams becoming usable: today they’re possible but awkward; a year out they may be the default for non-trivial work.
The shape of the next IDE
One forward-looking observation worth calling out: the “IDE” as we know it is converging on something that isn’t a code editor anymore.
- The primary surface is a chat interface, not a code buffer.
- A live preview renders whatever the agent is building, usually in a browser, sometimes in a running app.
- A document viewer shows the spec and the artifacts; you rarely type directly into either.
- You talk, the agent builds, you see the result. Conversation is the primary surface.
VS Code today is a compromise: it pretends you’re still writing code. The real shape of the workflow looks more like an iTerm pane + browser preview + agent, with the code file as an output you read rather than something you author character-by-character.
Claude Code’s desktop and web apps are steps in this direction, but nobody has fully landed it yet. Several teams are trying. The interesting question is who gets there first and whether it stays open enough to build on.
What the curriculum probably needs to add or revise next
This course is a snapshot of what makes sense in 2026. It will age. The parts likely to age fastest:
- Install and configure: UI flows shift quickly; specific commands and flags change.
- Modes and abilities: feature names and shapes evolve (auto mode, plan mode, the permission model) as the harness improves.
- Infrastructure choices: the vendor landscape churns; specific recommendations go stale.
The parts most likely to age well:
- The thesis: concept as product, artifacts as outputs. This is orthogonal to tooling.
- Context discipline: finite window, compaction, what belongs where. This is physics.
- The patterns: rewinding, branching, synthesize, failures-as-investment, reflection. These are craft.
Read this course for the craft and principles. Read the docs for the specifics.
Resources
- Recent posts from Anthropic, OpenAI, and Google research blogs
- Simon Willison’s blog, one of the best running chronicles of the frontier
- Andrej Karpathy’s YouTube channel for longer-form explanations of what’s happening under the hood