101: The Landscape

Overview

This chapter asks the question the rest of the course builds on: where are we? AI coding tools now run from autocomplete to agents that read a codebase, write the code and run the tests. The tools are plentiful and a practice for working with them is scarce, and in the reading of the course's landscape survey, taken in April 2026, most developers use the tools at the autocomplete end. The six insights map that landscape, then turn to the numbers used to describe it and to what cheap code does to the person generating it. The course's own evidence comes from the full git histories of three projects: Intent, Lamplight and Laksa.

The chapter also sets out the course map. The first year, 1xx, is for readers new to agentic coding: one human and one agent. The second year, 2xx, gives the practitioner a method at project scale. The third year, 3xx, runs multi-agent fleets, and comes last because its entry condition is fluency with a single agent. The chapter previews the course's evaluation framework too, a self-assessment that is course design rather than a finding from the record: the first year closes in 104 by scoring its six single-agent dimensions, and 301 adds the seventh.

By the end of this chapter, attendees should be able to place their own use of AI coding tools on the agentic-assistive spectrum, install Claude Code and judge a first delegated task by its result rather than its report, tell a labelled estimate from a count that carries its ref, its date and its counting convention, and name the need behind the work they generate.

Setup

Prerequisites: Software development experience in any language. No prior experience with AI coding tools is needed: Exercise 1 installs Claude Code.

Materials: A laptop with a terminal, and a small project you know well, under git and ideally with a test suite. Claude Code is installed during Exercise 1. An assistive tool such as Copilot or Cursor helps with Exercise 3 but is not required.

Time allocation: 90 minutes theory, 80 minutes practical.

Session Flow

ACI-001: The Agentic-Assistive Spectrum (15 minutes)

Key point: AI coding tools sit on a spectrum from assistive, where the AI suggests and you accept, to agentic, where you delegate and the AI executes. What moves a developer along it is readiness rather than tool capability, because agentic work needs a different mental model, and an agentic tool used with an assistive one gives assistive results. How many agents one human runs is a second axis, which the third year takes up.

Teaching notes: Open with a show of hands: who has used an AI coding tool for anything beyond autocomplete? Draw the four levels as the April 2026 survey set them out, with the developer's role at each: assistive, where the developer drives and accepts ghost text; conversational, where the developer directs a chat; collaborative, where the AI edits files and the developer reviews; and agentic, where the developer delegates a multi-step task and verifies the outcome, as with Claude Code. The placements are the survey's and have not been re-checked. Use the page's anti-pattern, the Micromanager, who installs Claude Code, approves every file read, rephrases each task as a single step, and concludes that it is "no better than Copilot". From the assistive end, the next step is conversational, not agentic. Then draw a second axis, for how many agents one human runs, and set out the course map on the two. The first year makes the move from suggestion to delegation with one agent, the second gives that practitioner a method at project scale, and the third moves along the second axis, starting in 301, once the agentic end of the first is familiar.

Transition: "The spectrum shows how far the tools reach. The next insight is about what is missing around them."

ACI-002: The Methodology Desert (15 minutes)

Key point: The agentic coding ecosystem has abundant tools and almost no reusable methodology for working with them over months. Judge a tool by the practice around it as well as by what it can do, and count what a framework ships rather than taking its description of itself.

Teaching notes: Separate tool documentation from methodology. Each tool's documentation says where project rules go, whether CLAUDE.md, .cursorrules or a conventions file for Aider. When the page was first written, in April 2026, a search for a methodology to go with them found almost nothing: no standard for evolving those rules, no session lifecycle, and no verification beyond "run the tests". The survey's lists have not been re-checked since. Do not read ChatDev or MetaGPT as covering the multi-agent case: in them agents coordinate with each other and the human stays outside the loop, while a human directing a fleet had no practitioner framework either. Then count one framework instead of describing it. Intent, the framework behind this course's projects, ships 68 rules, 23 skills (packaged instructions the agent loads for a kind of task) and 9 subagents (separate agents a session can hand work to), 6 of them critics that check code or prose against the rules (Intent@4c172d260, 2026-09-14). These are counts of what ships, not evidence that any of it works, and the course leaves the comparison with other frameworks to the reader. Close on the smallest framework one developer can keep: a rules file revised whenever the agent repeats a mistake, the same start and finish for every session, and a check on each report.

Transition: "Without a practice around them, much of what the tools can do goes unused. The next insight estimates how much, and says how far to trust the estimate."

ACI-003: The Ten Percent Problem (15 minutes)

Key point: Most developers use AI coding tools at the autocomplete level, which the April 2026 survey estimated at roughly a tenth of what the tools can do. Every share and multiplier behind that name is an estimate: label it as one, and ask any such figure for its source, its date and how it was counted.

Teaching notes: Put the survey's estimated shares of users on the board: autocomplete about 80%, chat about 15%, multi-file editing about 4%, and autonomous multi-step work about 1%. Say plainly that no study or repository behind this course measured those shares, and that the insight's name is itself an estimate. Give the survey's reasons the gap persists: the entry point shapes the ceiling, few resources teach the step up to delegation, and organisations measure adoption, meaning how many developers have the tool, rather than utilisation, meaning how well they use it. Beside the estimates, show a count the course can re-run: Lamplight, an Elixir application and one of the course's three projects, holds 11,894 commits as of 2026-09-14 (Lamplight@93ac51c68, 2026-09-14). The count names its commit and its date, and it describes activity, not productivity. Then ask the room what each enterprise figure measured. McKinsey estimated that the direct impact of generative AI on software engineering productivity could equal 20–45% of current annual spending on the function, a modelled estimate of value rather than a measured gain. GitHub reported that developers in a controlled experiment finished a programming task 55% faster with Copilot: coding speed on one task, not software delivery. The course checked each study's date and headline figure, and nothing more.

Transition: "The tools' own documentation does not close that gap. The next insight is about why it cannot."

ACI-004: The Tool-Practice Gap (15 minutes)

Key point: Anthropic tells you how to use CLAUDE.md and not how to keep that file true across months of real work, and no AI coding tool vendor documents the practice around its tool. Documentation gets a developer to adoption, and the practice that follows takes months and depends on the project.

Teaching notes: Set Anthropic's guidance, as the April 2026 survey read it, beside practice as of 2026-09-14. "Put project rules here" became a guidance file kept as an index of one-line rules under a load budget, each tracing to an incident, which 102 teaches. "Run tests" became believing a green only once you know it could have gone red, which 104 teaches. The session's lifecycle, which the guidance did not address, became the same start at every session and after every compaction, with the handoff written before a compaction, which 103 teaches. The guidance side has not been re-checked. A vendor cannot write this practice in advance, because it depends on the project, the team and what the agent got wrong last month. Tell the anti-pattern, RTFM: developers told to read the documentation configure the tool correctly and still use it at the autocomplete level. Close by previewing the course's evaluation framework, which profiles a tool and the practice around it on seven dimensions, each scored from 1 to 5: autonomy level, context engineering support, session continuity, verification and review, team and enterprise readiness, methodology ecosystem, and multi-agent coordination. The profile matters more than the total. The framework is course design rather than a finding, a structured self-assessment and not evidence. The first year closes in 104 by scoring the six single-agent dimensions, and 301 adds the seventh.

Transition: "A profile describes a practice on several dimensions at once. The next insight is about the single number people reach for instead."

ACI-005: The Productivity Claim (15 minutes)

Key point: Expect agentic development to speed work up unevenly: mechanical work can accelerate sharply, creative work much less, and scope-sensitive work can go backwards. Report productivity as that distribution, and print every count with its ref, its date and its counting convention, because a count without them cannot be reproduced, reconciled or refreshed.

Teaching notes: Begin with the page's own April table, which gave Lamplight 1,505 commits and 97 completed steel threads, the project's units of work, with no ref and no date. Both figures were right, and each reproduces exactly at Lamplight@a442c8d9b (2026-04-06). Twenty-three days later the repository held 1,968 commits (Lamplight@6c5657944, 2026-04-29). Without a date, a correct snapshot and a wrong figure look the same. The Laksa row was worse, because it compared two programs: 110 commits of an Elixir command-line tool, ending at Laksa@f1f2eb34b (2023-11-15), then, after twenty months without a commit, a port of Postcard, a Rails website builder, to Elixir (Laksa@8433c8d39, 2025-07-14). A commit can also change its unit: once Lamplight set up a whiteboard for concurrent agent sessions (Lamplight@fce7c0b7d, 2026-05-18), its commits carried a fleet's coordination as well as its code. Then give the task-type bands as the estimates from experience they are: 10–50x for systematic audits and refactoring, 5–10x for boilerplate, 2–3x for standard features, about 1x for architecture and scoping, and below 1x for scope-sensitive work. Keep every band off any single instance. A Lamplight audit that checked about 734 files against 15 coding rules and remediated all 389 violations it found had no multiplier measured for it (Lamplight@f7ed808b5, 2026-03-04).

Transition: "Velocity counts the output, not whether it was the right output. The last insight is about what cheap output does to the person producing it."

ACI-006: The Generation Trap (15 minutes)

Key point: When writing code costs almost nothing, the natural response is to write more of it, and that response is wrong, because agentic coding moves value from implementation to judgement about why and what to build. The scarce skill is knowing what is worth building, so measure progress by what was validated, not by volume.

Teaching notes: Tell MeetZaya's story and mark it as illustrative: the project sits outside the repositories this course re-measures, and the account is the developer's own. It was a career co-pilot tool built with Elixir and Phoenix over 12 months by an experienced engineer who had recently started using agentic tools, and who called generating working code at speed "pathologically addictive". By the developer's own judgement most of the early code was poor: it worked in isolation, did not compose into a coherent system, and solved problems nobody had validated. By the time the developer corrected course, the project faced strategic problems no code could fix: a co-founder's departure, a positioning problem, no data moat and too little funding. It was cancelled with 1,536 commits across 65 steel threads, 52 of them completed, and working software ready to test with real users. Then the page's analogy: a DJ who fills a stadium has the same tracks as everyone else, and what fills the stadium is taste. The tracks are free, and the taste is not. Present the trap as the warning for a developer's first weeks with an agent, and leave the room with the page's signal: if you feel productive but cannot name the user problem you solved this week, you are in the trap. The second year takes up the trap's second form, a method that makes verification cheap and then generates checks about checks.

Individual pathway notes

Give ACI-001 and ACI-006 the most time. Ask each attendee where they sit on the spectrum and what one step up would look like this week, remembering that from the assistive end the next step is conversational. Then leave them with two habits one developer can start today: write every count with its commit, its date and its counting convention, and at the end of each week name the need behind the work the agent produced.

Enterprise pathway notes

Give ACI-003 and ACI-005 the most time. A team with a licence for every developer can show high engagement on a usage dashboard while almost every accepted suggestion is a line or two of code and delegation goes unused. Track utilisation beside adoption by asking developers when they last delegated a multi-step task. Report return on investment as a distribution by task type, with each band labelled as an estimate or tied to the tasks and dates it was measured on, and budget as much time for the practice as for choosing the tool. This is advice rather than evidence: the course's record comes from one practitioner's projects, not from a team.

Connections

To 102: This chapter ends on judgement about what is worth building. 102 starts the practice with what the agent is given before it starts: delegation as a written brief with stated constraints and a checked outcome (ACI-007: The Delegation Model), and the guidance file the agent loads at every session start (ACI-010: The Guidance File).

Ahead to 301: The spectrum in this chapter measures how much one agent does on its own. The second axis, how many agents one human runs at once, is taught in the third year, starting at 301: The Step Up with ACI-055: SAAC to MAAC: The Step Up, where the evaluation framework also gains its seventh dimension.