A talk for engineers · 29 minutes · source: aihero.dev

The new starter
with no memory.

The basics of agentic development, from Matt Pocock’s AI Hero. What an agent is, why it fails, and the flow that keeps it honest.

Before we start · two things to hold

This is not a machine learning topic. The agent writes the code. You decide what is worth building, and whether what came back is right. Your engineering judgement is the asset. It transfers directly.

“AI is not a super-powered developer. It’s a new starter with no memory.”

aihero.dev/how-to-make-codebases-ai-agents-love

No memory between sessions. Every session starts empty.

The mechanism · start here

“An agent, though, doesn’t use predetermined steps. It calls an LLM and gives it a bunch of tools, different options of things that it can do next.”

aihero.dev/what-is-an-agent

A tool is one action it can ask for: read a file, run a command.

❯ the test in sum.test.js fails - find out why and fix it
     ↑ you typed this once
● Bash(ls -la …/demo)              ← the model asks for a tool
  ⎿  total 16                      ← the harness runs it
     -rw-r--r--  1 root root  164 Aug 29 18:54 sum.js
     -rw-r--r--  1 root root  198 Aug 29 18:54 sum.test.js
● Read(…/sum.js)
  ⎿  Read 10 lines
● Update(…/sum.js)
 3 -  for (let i = 0; i < numbers.length - 1; i++) {
 3 +  for (let i = 0; i < numbers.length; i++) {
 Do you want to make this edit to sum.js?
 ❯ 1. Yes   2. Yes, and allow …   3. No
     ↑ it stops and asks

sum.js                                           ← the code
function sum(numbers) {
  let total = 0;
  for (let i = 0; i < numbers.length - 1; i++) {    ← the bug
    total += numbers[i];
  }
  return total;
}

sum.test.js                                      ← the test
test('sums all the numbers', () => {
  assert.equal(sum([1, 2, 3, 4]), 10);              ← must give 10
});

❯ node --test
not ok 1 - sums all the numbers           ← one word: not ok
  error: '6 == 10'
  expected: 10
  actual: 6                               ← it got 6, not 10
# pass 0
# fail 1

“When working with AI coding agents, especially those operating independently, you need feedback loops so the AI can verify its own work.”

aihero.dev/essential-ai-coding-feedback-loops-for-type-script-projects

● Bash(node --test 2>&1 | tail -20)
  ⎿  ok 1 - sums all the numbers            ← the same test, now passing
     # pass 1
     # fail 0

● Off-by-one in the loop: sum.js:3 stopped at numbers.length - 1,
  so it skipped the last item and returned 6. I changed the bound
  to numbers.length; the test passes.

“Tests and feedback loops are essential — they’re how your AI new starters know if their changes work.”

aihero.dev/how-to-make-codebases-ai-agents-love

“AI agents don’t get frustrated by repetition. When code fails type checking or tests, the agent simply tries again.”

aihero.dev/essential-ai-coding-feedback-loops-for-type-script-projects

“Poor feedback loops. The AI doesn’t receive feedback fast enough, so it doesn’t know if what it changed actually did what it intended.”

aihero.dev/how-to-make-codebases-ai-agents-love

The fact behind every failure you will hit this week

“An agent can only work with what’s in its context window.”

aihero.dev/plan-mode-introduction

“Agents often work in fresh context windows.”

aihero.dev/my-7-phases-of-ai-development

What is actually in the window

Measured in tokens — chunks of text. A thousand tokens is roughly 750 words. A full window is about a short novel.

“Same model, same harness — just more context.”

“The zones don’t track the context window limit. A session can be deep in the dumb zone with most of the window still free”

aihero.dev/ai-coding-dictionary/smart-zone

Why quality falls · the attention budget

Attention is a fixed budget, spread across everything in the window.

“Like humans, who have limited working memory capacity, LLMs have an “attention budget” that they draw on when parsing large volumes of context.”

anthropic.com/engineering/effective-context-engineering-for-ai-agents

“Use /clear frequently between tasks to reset the context window entirely”

code.claude.com/docs/en/best-practices

/clear is a slash command. You type it into the agent.

When it drifts from what you told it, do not paste in more documentation. That makes it worse. Clear it, and start again.

• Kitchen sink“You start with one task, then ask Claude something unrelated… Context is full of irrelevant information.”
• Corrections“…you correct it, it’s still wrong, you correct again. Context is polluted with failed approaches.”
• Exploration“You ask Claude to ‘investigate’ something without scoping it. Claude reads hundreds of files, filling the context.”

“If you’ve corrected Claude more than twice on the same issue in one session… Run /clear and start fresh with a more specific prompt that incorporates what you learned.”

code.claude.com/docs/en/best-practices

No memory · so where do you write it down?

“CLAUDE.md is a special file that Claude reads at the start of every conversation… This gives Claude persistent context it can’t infer from code alone.”

code.claude.com/docs/en/best-practices

“For each line, ask: ‘Would removing this cause Claude to make mistakes?’ If not, cut it. Bloated CLAUDE.md files cause Claude to ignore your actual instructions!”

code.claude.com/docs/en/best-practices

When a fact becomes a procedure · move it out of CLAUDE.md

“If an entry is a multi-step procedure or only matters for one part of the codebase, move it to a skill… instead.”

code.claude.com/docs/en/memory

“Only the name and description sit in context by default. When the agent’s task matches, it loads the rest. Until then, the skill takes up almost no room — a sentence or two of tokens, however large its full instructions are.”

aihero.dev/ai-coding-dictionary/skill

The session ends. Everything you agreed dies with it.

“The spec exists because context windows end.”

aihero.dev/skills-to-spec

“the tickets are disposable and the spec isn’t” — a session dies in the dumb zone, and the spec is what survives into the next window.

• /grill-with-docs“Get interviewed about a plan, and record the decisions.”
• /to-spec“Turn an agreed conversation into a written spec.”
• /to-tickets“Split a spec into small tickets an agent can build.”
• /implement“Build a finished spec into code, test-first.” — /tdd is the engine inside this step.
• /code-review“Review a diff against your standards and against the spec.”

You write the brief once. Every later session reads it.

“You wouldn’t chuck vague requirements at a human colleague and expect good results. Without plan mode, that’s exactly what you’re doing to an AI.”

aihero.dev/plan-mode-introduction

“In plan mode, you iterate with the agent on what you want to build before any code gets written.”

aihero.dev/plan-mode-introduction

• Types“TypeScript is essentially free feedback for your AI. Use it over JavaScript.”
• Tests“Basic unit tests covering core functionality help keep the AI on track.”
• Hooks“If any step fails, the commit is blocked and the AI gets an error message.”
• Format“This runs Prettier on all staged files and automatically restages them.”

aihero.dev/essential-ai-coding-feedback-loops-for-type-script-projects

“Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one. And finally, if a question can be answered by exploring the code base, explore the code base instead.”

aihero.dev/5-agent-skills-i-use-every-day

“Skills don’t have to be long to be impactful. You just need to choose the right words at the right time.”

aihero.dev/5-agent-skills-i-use-every-day

• MondayOne real ticket, in plan mode. Shift+Tab. No code until the plan is right.
• TuesdayRun /grill-with-docs on the next ticket. Keep the file it writes.
• This weekOne feedback loop the agent can fail against. A typecheck, or a pre-commit hook.
• Every timeRead the diff. You are still the engineer. Nobody else is reviewing this.

Thank you

“You have access to a fleet of middling to good engineers that you can deploy at any time. But these engineers have a critical flaw: they have no memory. They don’t remember things they’ve done before.”

aihero.dev/5-agent-skills-i-use-every-day

No memory.
So write it down.