Moving upstream is not a career strategy
The full argument · 4 minutes
Architecture is a temporary altitude
The standard advice to senior engineers is to stop competing with AI at coding and move upstream into architecture, planning, and review.
It is comforting. It is incomplete.
METR's task-horizon research estimates that the length of software and reasoning tasks agents can complete at 50 percent reliability has doubled about every seven months.
Evidence:Task horizons
The benchmark is not the whole workplace. Its direction still makes one conclusion unsafe: there is no permanent altitude at which model capability stops climbing.
Moving upstream is useful work. It is not a career moat.
Architecture artifacts can be generated. Review can be partially automated. Planning horizons will expand.
The durable advantage is ownership of a closed feedback loop: choose a valuable problem, reveal the constraints the repository cannot reveal, define acceptable evidence, observe production behavior, and correct the system.
AI can execute more of that loop over time. Somebody must still make it correspond to reality.
Seniority matters where the task touches reality
Anthropic analyzed roughly 400,000 Claude Code sessions and found a recognizable division of labor: people made about 70 percent of planning decisions while the agent made about 80 percent of execution decisions.
Evidence:Expertise study
Task-specific expertise correlated with better outcomes, but most of the gain appeared between novice and intermediate knowledge. Deeper mastery added less.
Because this is vendor research based on model-classified transcripts, it is suggestive rather than causal. Still, it points away from generic tenure and toward knowledge of this customer, this system, and this failure mode.
Our discovery is that experience only compounds when the machine and the next engineer can use it.
A private intuition catches one mistake. The same intuition expressed as a Context Brief, invariant, architecture test, review rubric, rollout limit, or production alert can catch an entire class of mistakes.
Seniority becomes leverage when tacit judgment turns into executable constraints and observable feedback.
The context brief is a unit of leverage
A useful brief is not a biography of the repository.
It contains task-specific intent, non-goals, hidden constraints, known failure cases, and the evidence required for acceptance.
Evidence:Context-file studyHarness practice
Stable rules belong in AGENTS.md. Discoverable facts belong in the code. Proof belongs in CI and telemetry.
A context line earns its place only if removing it changes a decision.
That boundary matters because more context is not automatically better.
A controlled study across 438 repository tasks found that generated context files slightly reduced average success while increasing cost by roughly one-fifth. Human-authored context showed a small, statistically uncertain gain.
Evidence:Context-file study
The practical lesson is not to delete documentation. It is to reserve scarce context for information the agent cannot cheaply rediscover — and that changes the shape of a correct solution.
Run the two-by-two test
Select 20 to 30 historical tasks with hidden acceptance checks and blind maintainer review.
Run four conditions: an economical model and a frontier model, each given either the original ticket or a senior-authored Context Brief.
Hold tools and time constant. Measure first-pass acceptance, violated invariants, severe mistakes, review minutes, total cost, and rework.
The theory predicts that on work with implicit product or repository constraints, the expert brief will improve accepted outcomes more than the model upgrade.
Measure judgment leverage as verified changes per hour of senior attention.
If the frontier upgrade consistently dominates the brief, or the brief adds review without improving outcomes, the theory narrows: the task may be primarily capability-bound, or the senior has documented information that was already discoverable.
Protect the skill required to distrust the machine
Delegation creates an oversight paradox.
The more implementation and debugging an engineer delegates, the less practice they get in the skills needed to catch an agent's mistakes.
In a small randomized Anthropic study, the AI-assisted group scored 50 percent on a follow-up quiz versus 67 percent without AI, with the largest gap in debugging.
Evidence:Skill-formation RCT
The sample was only 52 people and measured immediate learning — so it should guide an experiment rather than a ban.
Reserve part of the AI dividend for comprehension: require a failure model before merge, ask for competing explanations, run occasional AI-off debugging drills, and use teach-back when the technology is unfamiliar.
Test delegation-only, generation-plus-teach-back, and manual groups on seeded faults two weeks later. If teach-back does not improve delayed fault detection, remove the ritual.
Senior engineers stay ahead not by refusing automation, but by continually proving that they can still verify it.
Research trail
Primary and independent sources support the factual claims. The Medium reading maps the public conversation; the argument above is our synthesis.