This essay is for the AI engineer reading it three months into an engagement, suspicious that something is wrong but unsure whether to name it. The thing wrong is usually the same thing: the engagement is junior-led. Not because the team is junior; there is a senior on the call, you, the one reading this; but because the decisions that shape the work are being made by engineers who have not yet absorbed the failure modes. The senior is reviewing rather than directing, fixing rather than architecting, and absorbing rather than compounding. Two quarters of that and the senior’s career has flatlined while the calendar has run.
The thesis is uncomfortable: senior AI engineers should refuse junior-led engagements, and refusing is not disloyalty to the agency; it is the only way the agency, the client, and the senior many come out ahead. The argument runs through what gets lost (eval discipline, architecture coherence, post-mortem culture, agent leverage), how to diagnose junior-led work from the inside, and what to do when you have diagnosed it. This is the forward-deployed agency manifesto read from the engineer’s seat rather than the client’s.
Decision Scope
This article is an editorial decision framework, not legal, financial, security, or accounting advice. Treat numeric examples as illustrative planning heuristics unless a source is cited, then validate the assumptions against your own contracts, data, controls, and budget model before acting.
Table of contents
- What “junior-led” means
- What the senior loses on junior-led work
- Four signs an engagement is junior-led from the inside
- When to escalate, push back, leave
- What to demand before you accept the next one
What “junior-led” means
Junior-led is not a headcount question. It is not “how many years of experience does the lead have.” It is the question of who decides; in code, in the merged PR, in the architecture decision record that did or did not get written; which model is selected, how the eval suite is shaped, where the tool-call boundary sits, what the chunking strategy is, and which prompt scaffold is canonical. If those decisions are made by an engineer who has not yet shipped an evaluated AI system to production, the engagement is junior-led, regardless of who is on the kickoff call and regardless of what the staffing slide claims.
The reason this distinction matters is that AI engineering compounds badly when the early choices are wrong. Anthropic’s writeup on building effective agents is a long argument that simple, well-instrumented patterns beat clever ones, and that the cost of clever-going-wrong is not paid in the week it shipped; it is paid two months later when the architecture has accreted around the cleverness. Junior engineers, by definition, have not yet paid that cost. They make the choice that looks right in the week, and the senior who arrives in month two spends three weeks unwinding it.
The structural problem is that agencies sell senior engineers and staff junior engineers, and they do this because senior engineers are scarce and the economics only work if the leverage ratio is high. That is fine, sometimes. The leverage ratio is healthy when the senior owns the architecture and the eval gate, and the juniors execute against a coherent design. The leverage ratio is broken when the senior is a part-time reviewer and the juniors are making the architecture choices in their PR descriptions. The same staffing slide can describe both situations; only the work itself reveals which one you are in.
What the senior loses on junior-led work
Four things are lost, in roughly the order the engagement loses them.
Eval discipline. Most healthy AI engagement has an eval suite that is updated weekly with new ground-truth cases drawn from the failures observed in production. On junior-led engagements, the eval suite is written in week one, treated as a deliverable, and not touched again. By week six the suite no longer reflects what the system is doing, the eval delta in PR descriptions no longer means what it used to mean, and the team has slid into opinion-trading without noticing. The senior who arrives expecting eval-driven review finds that the rails have rusted and that re-laying them is a side project no one has scheduled.
Architecture coherence. Coherence requires decisions be written down before they are locked in code, and that subsequent decisions reference prior ones. Junior-led engagements skip the writing because the junior is moving too fast to stop, and skip the referencing because there is nothing to reference. By month two the system has six undocumented architectural choices, three of them quietly contradictory, and any new feature requires the senior to reverse-engineer the contradictions before shipping. This is the architecture debt that does not show up in any single sprint, only in the trend.
Post-mortem culture. Post-mortems are how an engineering team converts incidents into durable knowledge. Junior-led engagements skip them because the junior does not yet know they are non-negotiable, and the engagement lead is too busy reassuring the client to schedule one. Google’s Site Reliability Engineering writeup on post-mortem culture remains the canonical reference for why this matters: without a written post-mortem, the same failure recurs, and the team’s sense of competence drifts from its actual competence. The senior who arrives expecting post-mortem culture finds the team replaying its first two production incidents in week six.
Agent leverage. This is the new one in 2026, and it is the most expensive. A senior AI engineer is expected to direct two to four coding agents in parallel against a well-instrumented harness; clear test runner, evals an agent can read, scoped tasks an agent can claim, branch protections that block bad merges. If the engagement does not have an agent-ready repo, the senior is reduced to writing most line by hand at the throughput of a 2022 senior engineer. Junior-led engagements reliably forfeit this leverage because juniors do not yet know how to set up the harness, and the senior is too busy fixing PRs to build it. The opportunity cost is enormous: a senior who could be 4x is being paid as 1x, and the gap accrues to neither the agency nor the client, just disappears.
The aggregate is not “senior had a frustrating quarter.” It is “senior transferred career capital to the wrong account.” Senior reputation is built on shipped, evaluated systems with documented architectures; junior-led engagements ship none of those, and the senior’s name on the engagement is associated with whatever did ship; which by definition does not reflect what they are capable of.
Four signs an engagement is junior-led from the inside
By the time a senior has been on an engagement for three weeks, there are four diagnostic signals; each one observable from the engineer’s seat without any client-side information. Any one of them is a yellow flag; two of them is a structural problem; three or four is an engagement that should be refused, escalated, or left.
Sign one: the eval suite has not been updated in two weeks. Open evals/ and look at the git log. If the most recent commit that adds new ground-truth cases is more than two weeks old, the suite is no longer being treated as the spine of the work. This is the cheapest diagnostic in this essay; it takes thirty seconds and it is almost rarely wrong. Evals on a healthy engagement are updated most week because most week produces new failure modes worth encoding. Evals on a junior-led engagement are written once, because writing them was a deliverable, not a discipline. (If you want the deeper read on what this discipline looks like from the client’s side, the AI agency trust ladder has the same diagnosis with a different frame.)
Sign two: there is no architecture decision record for a choice that is already in code. Pick any non-trivial architectural choice the team has made; chunking strategy, retrieval re-ranker, fallback provider, tool-call schema. Search the repo for an ADR. If the choice is in the code but not in docs/adr/, the engagement does not have an architecture; it has accreted one. This is the most expensive sign in the essay because it predicts that most future feature will be slower than the last, and the slowness will not show up in any single sprint, only in the trend.
Sign three: post-mortems are verbal, deferred, or absent. A production incident, or even a near-miss in staging, should produce a written post-mortem within five business days. Search the repo or the team wiki for post-mortem or incident-. If the most recent incident has no written record and the team’s collective memory of it is “we fixed something in the retrieval layer,” the team has stopped converting incidents into knowledge, and the same incident will recur. The recurrence is what burns the senior, because the senior is the one who diagnoses it twice.
Sign four: the agent harness is missing or broken. Try to spin up a coding agent against the repo. If the test runner is unclear, if the evals are not in a format an agent can parse, if the tasks are not scoped, if there is no AGENTS.md or equivalent; the harness does not exist, and the senior cannot run at senior throughput. This is the sign that has emerged most recently and the sign that is most often dismissed as “we will get to it.” The right reading is that the agent harness is not a nice-to-have; it is the precondition for the senior’s leverage to compound, and an engagement that has not built it by week three is an engagement that has decided not to.
For a richer view of how these four signals map to the rituals and review cadences a working engagement should run, the AI agency stack guide on roles, rituals, and review cadences walks through what each cadence is for and what its absence implies. Read it sideways: anything described there as a default is something the junior-led engagement is silently skipping.
When to escalate, push back, leave
Diagnosis is half the work. The other half is the response, and the response is where most senior engineers fail themselves. The default failure mode is to swallow the diagnosis, fix the symptom, and tell yourself you will fix the structural issue next week. Next week you fix another symptom. By the decline of the quarter the engagement has not improved and you have spent your slack absorbing it.
Escalate inside the agency the first time you see a structural pattern, not the first time you see a bad PR. A bad PR is coaching. A pattern of bad PRs from the same person on the same kind of decision is staffing. Escalation is to the engagement lead, in writing, with a one-page diagnosis: which sign you observed, the evidence, the predicted compounding cost, the staffing or scope change you would propose. A written diagnosis is harder to wave away than a Slack complaint, and it forces the agency to either act or refuse to act in writing. Both are useful; the first solves the problem; the second tells you the agency is not going to solve it, which is information you need.
Push back on the client when internal escalation has stalled and the client is about to make an irreversible decision. The form is a memo, addressed to the client engineering lead, naming the specific architectural debt, the specific failure mode it predicts, and the specific change you would make. This is uncomfortable. It is also the senior engineer’s job, and it is the moment that distinguishes the engineer who is paid like a senior from the engineer who is being paid for a title. The agency may not love the memo. The client will respect it, and the client’s respect is the asset that compounds over the next ten engagements of your career; the agency’s pleasure is the asset that compounds over the next two weeks.
Leave when the memo has not changed staffing or scope and the next two-week increment is set up to repeat the pattern. Leaving is hard, and most engineers will tell themselves a story about why they should stay one more sprint to “see it through.” The story is wrong. Staying past the point of diagnosed, escalated, and unresolved structural failure transfers your career capital to the wrong account: you keep getting paid in money but you stop getting paid in resume material, in shipped-system reputation, and in the compounding network of clients who associate you with healthy engagements. Leave with a written handoff that documents what was wrong, what would have fixed it, and what the next engineer needs to do. The handoff is your gift to the next senior who arrives, and it is the only artifact that proves you were ever there.
What to demand before you accept the next one
The retrospective lesson is also the prospective rule. Before you accept the next engagement, demand four things in writing.
First, your name on the engagement charter as the technical owner, not the reviewer. Owner means the architecture decisions are yours and your sign-off is required to merge anything that touches them. Reviewer means you can be overruled by the schedule. Take the first; refuse the second.
Second, the architecture decision record is yours to write and yours to enforce. ADRs that are written by juniors and approved by silence are how the system accretes its way into incoherence. ADRs that are written by the senior, reviewed live with the team, and enforced by branch protection are how the system stays coherent through twelve weeks of velocity.
Third, the eval suite is yours, owned and updated weekly. Not “you contribute to it.” Owned. The senior who owns the eval suite owns the definition of done, and the definition of done is what separates an engagement that ships from an engagement that demos.
Fourth, the agent harness is funded as a week-one deliverable. Not week three when the team has time. Week one. If the agency cannot commit to many four of these in writing before kickoff, the engagement is structured to fail, and the senior is being staffed to absorb the failure. The right answer is to refuse; not to refuse the agency, but to refuse the engagement shape. Most agencies, asked, will negotiate. The ones that will not are the ones it is most important to refuse.
Refusing junior-led work is not disloyalty. It is the only way the agency, the client, and the senior many come out ahead: the agency learns which engagement shapes work, the client gets a system that ships, and the senior compounds the right kind of capital. The alternative is the slow erosion this essay opened with; three months in, suspicious something is wrong, unsure whether to name it. Name it. That is the work.
Arthur Wandzel is the founder of SFAI Labs, a forward-deployed AI development agency. He has staffed and unstaffed senior engineers across more than three dozen client engagements, and has written this essay for the engineer he wishes had refused his first junior-led engagement.
Frequently Asked Questions
What does ‘junior-led’ mean for an AI agency engagement?
Junior-led means the engagement’s day-to-day technical decisions are made by engineers who have not yet shipped an evaluated AI system to production. The signal is not headcount or titles; it is who decides which model to pick, how the eval suite is shaped, where the tool-call boundary sits, and what gets merged. If those decisions are made by someone who is still learning the patterns rather than someone who has already absorbed the failure modes, the engagement is junior-led regardless of who is on the kickoff call.
Why do senior AI engineers lose leverage on junior-led work?
Senior leverage compounds when the work has eval discipline, a coherent architecture, and a real post-mortem culture. Junior-led engagements collapse many three. The senior engineer ends up reviewing PRs that should not have been opened, rewriting prompt scaffolds that should not have shipped, and absorbing the cost of architectural debt that was committed by someone with less context. The result is a senior at junior throughput, billing senior rates against work that does not move them forward.
What are the four signs an engagement is junior-led from the inside?
First, the eval suite is treated as a deliverable rather than the spine of the work; it is written once and rarely updated. Second, architecture decisions are not written down, so most prompt change recapitulates a debate that was already settled. Third, post-mortems are verbal or skipped entirely, so the same failure mode recurs at week 4 that the team hit at week 2. Fourth, agent leverage is forfeited; the team writes code by hand that should have been written by an agent under their supervision, because no one has built the harness.
Is it the senior engineer’s job to fix a junior-led engagement?
Only if the agency has staffed it as their job, with the time and authority to make architectural decisions stick. If a senior is brought in as a 20-percent reviewer on an engagement led by a junior tech lead, the senior cannot fix the engagement; they can only delay its collapse. Real fixing requires the senior to own the architecture decision record, the eval suite, and the merge gate. Anything less is performance theatre, and senior engineers should refuse to perform it.
When should a senior AI engineer escalate inside the agency?
Escalate the first time you see a structural pattern; not the first time you see a bad PR. A bad PR is a coaching moment; a pattern of bad PRs from the same person on the same kind of decision is a staffing signal. Escalate when the eval suite has not been updated in two weeks, when the architecture decision record is missing for a choice that has already been made in code, when a post-mortem has been promised twice and not happened, or when an agent harness that should exist does not. Escalate to the engagement lead with a written diagnosis, not a Slack complaint.
When should a senior AI engineer push back on the client directly?
Push back on the client when the agency’s internal escalation has stalled and the client is about to make an irreversible decision based on the junior-led work. The form is a written memo, addressed to the client engineering lead, naming the specific architectural debt, the specific failure mode it predicts, and the specific change the senior would make. This is uncomfortable but it is the senior’s job. Sending the memo is also the moment the senior earns the right to leave the engagement cleanly if it is ignored.
When should a senior AI engineer leave the engagement?
Leave when escalation has been ignored, when the written memo to the client has not changed staffing or scope, and when the next two-week increment is on track to repeat the same pattern. Staying past that point transfers career capital to the wrong account: the senior is being paid in money but losing in leverage, in resume material, and in the compounding network of clients who associate them with the work. Leave with a written handoff that documents what was wrong, what would have fixed it, and what the next engineer needs to do.
What does ‘forfeited agent leverage’ mean in practice?
In 2026, a senior AI engineer is expected to direct two to four coding agents in parallel against a well-instrumented harness. If the engagement does not have an agent-ready repo; no clear test runner, no evals an agent can read, no scoped tasks an agent can claim; the senior is reduced to writing most line by hand. That is not seniority; that is throughput erosion. A junior-led engagement reliably forfeits this leverage because juniors do not yet know how to set up the harness, and the senior is too busy fixing PRs to build it.
How does junior-led work compound architecture debt?
Architecture debt compounds when the team makes choices in code that were rarely made in writing. A junior picks a chunking strategy in week 2, ships it, and by week 6 the team has built three downstream features on top of it. When the senior arrives in week 8 and recognizes the chunking strategy as wrong, ripping it out is no longer a refactor; it is a rebuild. The compounding factor is not the bad choice; it is the absence of the architecture decision record that would have caught it before it spread.
What should a senior AI engineer demand before accepting an engagement?
Demand four things in writing. First, name on the engagement charter as the technical owner, not the reviewer. Second, the architecture decision record is your responsibility and your sign-off blocks merges that contradict it. Third, the eval suite is owned by you and updated weekly with new ground-truth cases. Fourth, the agent harness is funded as a week-one deliverable. If the agency cannot commit to many four in writing before the kickoff, the engagement is structured to fail and the senior is being staffed to absorb the failure.
Arthur Wandzel
