Build a workflow you can defend
Coordinate small tasks, measure actual value, and show engineering judgment in your portfolio.
BY THE END, YOU CAN
- Split work while preserving clear integration ownership
- Evaluate AI assistance with quality and total effort measures
- Present a portfolio change with transparent evidence of judgment
Diagram description: Agent: API → Integration owner; Agent: tests → Integration owner; Agent: docs → Integration owner; Integration owner → Merged change
Parallel work still needs an integration owner
Give separate assistants independent tasks with explicit interfaces: one can inspect Relay’s retry contract while another reviews test gaps.
Avoid assigning two agents overlapping edits without coordination. Specify the deliverable, permitted scope, and completion evidence for each task. Integrate one coherent change at a time and rerun the relevant checks against the combined result.
Multiple assistants agreeing does not make a claim true, especially when they received the same mistaken premise. Use parallelism when independent work exceeds the cost of coordination and review; you remain responsible for the merged behavior.
Measure the whole job, including repair
For several comparable Relay tasks, record active work, waiting, review, rework, defects, and whether the result met the same acceptance standard.
Distinguish elapsed delivery time from your own effort when agents run concurrently. Small personal comparisons are useful observations, not controlled experiments. METR’s 2025 randomized study found a slowdown in its specific experienced-developer setting; its February 2026 update says newer estimates are unreliable, because developers who did not want to work without AI declined to take part, and time was hard to measure for those running several agents at once.
Neither establishes a universal effect for your work. Re-evaluate as tasks and tools change instead of treating perceived speed or generated lines as productivity.
Make your judgment visible to another engineer
A useful portfolio explains a problem, constraints, alternatives, and the evidence behind a decision.
Show a compact Relay specification, an architecture decision, a reviewed diff, a failure you reproduced, and test results. Describe where AI helped and which suggestion you rejected, including why. Rehearse explaining the system and making a small change without generated answers; follow an employer’s rules for assessments and confidential material.
For deeper practice, run an unfamiliar failure drill and revise the design. These artifacts demonstrate engineering work; completing a track does not guarantee employment or establish professional competence by itself.
COMMON FAILURE MODE
Measuring success by how much code an assistant produced, while ignoring review time, defects, and whether you can maintain the result.
TRY IT YOURSELF
Choose two similarly scoped changes and define a small personal workflow comparison. Record quality criteria and total review/rework effort before starting. State at least two reasons the result may not generalize.
Make the call in the laboratory.
The two-question checkpoint, your notes, and your saved progress are in the laboratory. Free, with no account.
Open in the laboratory ↗CONNECT IT TO RELAY
Export an AI delivery dossier: brief, specification, task and test map, review findings, final evidence, and a short reflection on assistance versus your own judgment. Explain one design decision aloud.