Context Engineering · Throw Away The Vibes
← → navigate · N notes · M menu · P read mode

DDD Melbourne 2026

Throw away the vibes.

Reliable results from AI coding agents don't come from prompting harder. They come from engineering the context you feed the model.

scroll to read — notes included P to present M contents dasith.me

The familiar story

The magic that evaporates.

Most of us feel an intoxicating high the first time we use an AI coding agent. Then we point it at a real codebase and the magic disappears.

Hello World

It works flawlessly

The toy example lands perfectly. It feels like the future has arrived.

Brownfield

The frustration hits

Point the same agent at a large business repo and the results fall apart. The tools aren't useless — we're feeding them the wrong context.

So the fix is not a better prompt. It's a better information environment.

The premise

Exceptional generators of code. Weak judges of architecture and context.

The instinct is to keep prompting across many turns until the agent eventually gets it right. The better goal is to fix the context so the first answer lands close to correct — that's context engineering, and it matters far more than the vibes.

Where the effort goes

Stop nudging. Curate the input.

Vibe coding leans on a hopeful loop: generate, notice it's wrong, nudge repeatedly until it converges. That loop is slow, it pollutes the conversation, and it rarely meets our bar.

Vibe coding

Correct after the fact

Let the agent generate, then prompt it again and again. Brilliant generator, weak judge — so the loop drags on.

Context engineering

Invest in the input

Curate exactly what the model sees before it writes a line, so the very first answer lands close to correct.

Reflex Once you internalize the generator–judge asymmetry, the job changes.

Part one

Why a long context window isn't enough.

The common reaction is to assume the fix is simply more context. In practice, that approach makes things worse.

The needle problem

A million tokens of noise.

Cramming thousands of files into a huge window doesn't help the model find the one detail that matters. The signal drowns.

The takeaway is not "less context" as a rule. It is just enough context, delivered at the exact moment it is needed.

Four failure modes of long context

How long context breaks down.

Poisoning

A hallucinated error early in a thread lingers and keeps corrupting everything generated after it.

Distraction

Attention weights vary across a long prompt, so the model misses crucial details buried in big blocks — the needle in a haystack.

Confusion

Superfluous, redundant information dilutes the signal and pulls the model toward irrelevant details.

Clash

Contradictory pieces of code in the same window leave the model unable to tell which one to trust.

The rule of thumb

Not less context. Just enough context — delivered at the exact moment it's needed.

Part two

The architecture of context engineering.

A distinct layer that sits on top of prompt engineering and beneath autonomous agents and opinionated engineering workflows.

Where it sits

A layer of its own.

Prompt engineering shapes the instruction. Context engineering shapes the information environment that instruction operates within — the layer everything else builds on.

Keeping context lean and relevant

Four architectural tactics.

Externalizing context

Move context out of volatile chat history into a shared space the human and agent edit as one source of truth.

Tool loadout

Restrict active tools, MCP servers, and skills to only what the problem needs, so the model isn't overwhelmed by options.

Compression

Summarize to condense with some loss, or compact by swapping full text for references the agent can retrieve later.

Isolation

Quarantine work into separate threads or sub-agent scopes so no single agent drowns in the broader orchestration.

Part three

Practical workflows.

Theory is only useful if it changes how you work on Monday morning. Two workflows turn these ideas into repeatable practice.

Workflow one

The Breadcrumb Protocol.

A lightweight human-and-agent pattern built around a single markdown scratchpad. The conversation is disposable. The file is durable.

1

Plan & break down

Human and agent co-author a task breakdown in a markdown file before any code is generated.

2

Iterate & log

As the agent executes, the file is continuously updated with state, decisions, and discoveries.

3

Quarantine failures

When a thread is polluted, abandon it. Keep the file, note why it failed, and feed it to a fresh session.

Workflow two

Research, Plan, Implement, Review.

The same principles scaled across a team. Microsoft's open-source hve-core structures development into a constrained, multi-step pipeline where each stage hands a clean, scoped context to the next.

Research Plan Implement Review

Isolation Each stage is the isolation tactic applied at the level of a whole development process.

Takeaways for 2026 and beyond

Practices worth adopting.

  • Watch your context threshold. Past roughly 60% capacity, compact it or start a fresh thread rather than pushing on.
  • Treat sub-agents as functions. Resist anthropomorphizing them — each is a discrete, tightly scoped step with clear inputs and outputs.
  • Shift the review left. Don't wait for the PR to review a wall of green text. Review incrementally as the work happens.
  • Master one harness. Stop chasing every new tool. Pick a core stack, learn its context limits, and get good at harness engineering.

Wrapping up

The vibes can go. The engineering stays.

Human expertise hasn't become less valuable. The role shifts from typing code toward scaffolding, steering, and the systematic curation of context.

Go deeper

The full write-up, the talk recording, and the workflows referenced here: