The fastest knowledge transfer between an AI agency and a client is not a documentation handoff; it is a four-to-six-week residency. A senior agency engineer sitting inside the client organization, attending their standups, reviewing their PRs, pairing on their evals, and going to their lunch. The output is not “improved collaboration”; it is a structurally different rate of knowledge transfer that documentation handoffs cannot match. This is the case for the rotating residency model, the conditions under which it works, the boundaries that keep it from becoming a BPO arrangement, and the pricing structure that makes it economically defensible.
The argument is empirical. Across roughly 24 engagements over the last 18 months, the ones that ran a residency in Q2 or Q3 had measurably better outcomes on three metrics: time-to-second-feature, eval coverage of long-tail cases, and post-engagement transition success. The mechanism is not mysterious; a senior engineer in the client building absorbs context that no documentation captures, and that context shows up in better architectural decisions, faster failure-mode diagnosis, and tighter eval cases. The residency is not a service offering; it is a delivery mechanism.
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 the residency is and is not
- The four reasons it outperforms documentation handoffs
- Q1 vs Q3 residency: when in the engagement does it fit
- Who rotates: lead engineer per quarter, eval owner per domain
- The boundaries that keep it from becoming BPO
- Pricing: how the residency is billed
- Logistics: travel, time zones, and remote residencies
- The objections and the counters
- When residency is the wrong fit
- FAQ
What the residency is and is not
A residency is a 4-to-6-week period during which a single named senior agency engineer works on-site (or in a structured remote-residency arrangement) embedded inside the client engineering organization. The engineer attends client standups, reviews client PRs, pairs with client engineers on eval design, and is co-located with the client domain expert for the duration. Their work artifacts ship into the client repo, not into the agency’s repo and then transferred.
A residency is not a placement, not a contract-to-hire arrangement, not a body-shopping engagement. The agency engineer remains an employee of the agency, billed at the agency’s rate, with the agency as employer of record. The deliverables of the residency are explicit and bounded; typically one major feature shipped, one eval-coverage expansion completed, and one knowledge-transfer artifact produced; and the residency ends on a defined date with a defined exit.
The residency also is not an offer to provide on-call coverage or to operate the system long-term. The model is structurally different from outsourced operation: the residency is a leverage mechanism for accelerating an engagement, not a substitute for the client owning the system.
The four reasons it outperforms documentation handoffs
1. Fastest knowledge transfer. Documentation captures explicit knowledge well and tacit knowledge poorly. The unwritten rules of an engineering organization; when to interrupt the on-call, which Slack channel handles which class of question, which Jira label is treated as P0 versus aspirational P0; are absorbed by being present. A residency engineer learns these in a week; a documentation handoff misses them entirely. The compound effect across a 90-day engagement is large.
2. Deepest data understanding. AI systems are dominated by data understanding. The eval cases that matter, the failure modes that recur, the long-tail inputs that surface in production; many of these are understood by sitting next to the people who handle them daily. An agency engineer who has watched the customer support team triage 40 ambiguous AI outputs over three weeks has a different mental model than one who reads a doc describing the same.
3. Real-time eval pairing. Eval design is a collaboration between an engineer who knows what the model can do and a domain expert who knows what the user needs. That collaboration is high-bandwidth and benefits from same-room presence. The eval cases that come out of a residency-paired session are sharper, fewer, and more representative than the ones that come out of an asynchronous Slack thread. For more on why eval design is the structural anchor of an engagement, see stop scoping AI projects in features, scope them in evaluations.
4. Builds relationships. This is the soft factor that the other three depend on. A residency engineer who has had lunch with the client team, attended their offsites, and helped them debug a non-AI issue at 4 PM on a Tuesday has a relationship surface that survives the eventual transition. When the post-residency handoff happens, the trust is already in place; when a stress event hits in month seven, the residency engineer is the person the client calls.
Q1 vs Q3 residency: when in the engagement does it fit
The residency is most powerful in Q3 of the quarterly mandate cadence. By Q3 the system is in production, the eval suite is mature, and the open question is expansion; adding the second feature, integrating the second data source, extending to the second user segment. Q3 is when the agency engineer’s deep client-context understanding has the highest leverage, and when the residency builds the relationship surface that supports a Q4 transition.
Q1 residencies are sometimes appropriate but more often premature. The Q1 work is establishing the eval baseline, shipping the first eval-gated PR, and writing the architecture decision record; work that can be done effectively from the agency office with the two-day kickoff providing the initial context download. A Q1 residency is right when the data understanding is the rate-limiting step (highly regulated domain, sensitive data, complex existing pipeline) and wrong when the engagement is in a familiar shape.
Q2 residencies are also possible but rarely the highest-leverage placement. Q2 hardening work; observability instrumentation, eval coverage expansion, cost reduction; is largely engineering work the agency can do well remotely. The exception is when the production failure modes are unfamiliar enough that a residency engineer learning them in real-time pays back the cost.
The structural shape: at most one residency per quarter, and the engagement should not run more than two residencies in 12 months. Three or more residencies in a year is no longer a residency model; it is an embedded-engineering arrangement under a different name.
Who rotates: lead engineer per quarter, eval owner per domain
The rotation is not arbitrary. The two roles that most benefit from a residency are the lead engineer and the eval owner. The lead engineer rotation places the senior engineer who owns the architecture in the client building for the architectural decisions of the quarter. The eval owner rotation places the engineer who maintains the eval suite next to the domain expert who provides ground-truth cases.
The lead engineer rotation is per-quarter. A different lead engineer can rotate each quarter, which has a structural advantage: it spreads the client-context understanding across the agency rather than concentrating it in one person, which prevents key-person risk on the agency side. The eval owner rotation is per-domain; when an engagement spans multiple domains (customer support, sales operations, content moderation), each domain has its own eval owner, and each does a residency when their domain is the active focus.
The agency engineer who does the residency is selected for fit, not for availability. The criteria: senior engineering experience (this is not a junior placement), strong written communication (the residency artifacts must be high quality), explicit alignment with the client’s working hours and culture, and willingness to travel or work the residency arrangement. Engineers who do not meet these criteria do not rotate.
The boundaries that keep it from becoming BPO
The residency model is structurally close to several arrangements it must not become. The boundaries below are what distinguish a residency from a placement, a body-shop, or a BPO.
The agency remains employer of record. The residency engineer is paid by the agency, has agency benefits, and reports to an agency manager. The client does not write paychecks, does not manage the engineer’s career, and does not have hiring authority over them.
The residency has a fixed end date. Most residency has a written start date and end date in the engagement charter. Extensions require an explicit re-scope, not a default rollover. Open-ended residencies become embedded arrangements and lose the structural advantages of the residency model.
The deliverables are bounded and named. Most residency has three named deliverables in writing; typically one feature shipped, one eval expansion completed, one knowledge-transfer artifact produced. If by the end date the deliverables are not done, the residency closes with a written gap analysis, not an extension.
The residency engineer does not become the on-call. The client on-call rotation does not include the residency engineer. The residency builds the muscles the client team needs to operate the system; it does not become the operation.
The residency does not displace client hiring. The residency is a knowledge-transfer mechanism, not a substitute for the client building their own AI engineering capability. A client who runs sequential residencies in lieu of hiring is signaling a structural mismatch; either the client should hire, or the engagement should transition to a different shape.
For the broader contract structure that includes residency provisions, see the AI agency annual contract is dead.
Pricing: how the residency is billed
The residency is priced as a fixed-fee deliverable inside the relevant quarterly mandate, not as time-and-materials. The fee covers the engineer’s time at the agency rate (typically 1.2-1.4x the engineer’s standard hourly rate to account for residency overhead), plus the named travel and accommodation costs at receipts, plus the residency artifacts delivered. The total fee is named at the start of the quarter; there is no per-hour billing during the residency.
Fixed-fee residency pricing aligns incentives. A time-and-materials residency creates an incentive for the residency to extend; a fixed-fee residency creates an incentive for the engineer to deliver the named artifacts and exit on schedule. The agency’s margin on a residency is not in the hours; it is in the price of the artifacts.
The residency premium over a non-residency week is typically 15-25%, reflecting the overhead of co-location and the opportunity cost to the agency of having a senior engineer outside the home office. Engagements that cannot defend that premium against expected outcomes should not run a residency; engagements that can are typically the ones where the residency outcomes are decisive.
For the broader pricing argument, see the AI agency pricing manifesto.
Logistics: travel, time zones, and remote residencies
The default residency is on-site at the client office. Travel and accommodation are billed at receipts; the agency does not mark up logistics. The on-site cadence is typically 4 days per week in the client office, 1 day remote for agency-internal work; a full 5 days on-site loses the agency-side context the engineer needs to maintain.
Remote residencies are a real variant, particularly for engagements where on-site is not practical (international clients, distributed client teams, regulated industries with access constraints). A remote residency works when three conditions hold: the residency engineer’s working hours overlap with the client’s by at least 6 hours, the client has a synchronous-communication culture (strong video presence in standups, willingness to pair on a call), and the deliverables can be unambiguously produced without same-room presence. Remote residencies typically have 80-90% of the leverage of on-site at 60-70% of the cost.
Hybrid residencies; two weeks on-site, two weeks remote; work well when the client is geographically distant and the team can absorb the context download in a focused on-site stretch followed by a remote-execution stretch.
The objections and the counters
“Residencies are expensive.” True at sticker price; the right comparison is residency cost versus the cost of a missed Q3 expansion or a botched Q4 transition. On engagements where the residency is well-fit, the residency cost is typically 6-10% of the quarter’s total fee and produces 20-30% better outcomes on the named metrics.
“We don’t have desk space.” Solvable. Hot-desks, conference room residencies, even open-office adjacencies many work. The point is presence in the building, not a permanent desk allocation.
“Our security policy doesn’t allow contractor on-site access.” Real and sometimes decisive. The fix is either a remote residency, a structured on-site visit cadence (one week per month rather than four contiguous), or a residency at the agency office where the client team rotates in instead. The model adapts to the constraint.
“We don’t want a contractor in our standups.” Reasonable concern, structurally addressed. The residency engineer is not a generic contractor; they are a named senior engineer doing 4-6 weeks of focused work with bounded deliverables. The client’s discomfort with “a contractor in standups” is usually a discomfort with body-shopping arrangements, which the residency model is structurally not.
“What if the residency engineer leaves the agency?” Handled by the rotation structure. Each quarter is a different lead engineer, and the artifacts the residency produces (eval expansions, architecture decisions, knowledge-transfer documents) live in the client repo. The relationship surface is broader than any single engineer.
When residency is the wrong fit
Not most engagement should run a residency. Three patterns indicate residency is the wrong fit.
The engagement is small. A residency on a 90-day engagement with a single feature in scope is overkill; the cost of relocation cannot be amortized against a small enough scope. Residencies fit engagements with $200K+ in quarterly fees and meaningful expansion ahead.
The client is hostile to embedded engineering. Some client cultures genuinely do not benefit from a residency; the team prefers asynchronous collaboration, the security model resists on-site contractors, the executive sponsorship is uncomfortable with the optics. Forcing a residency in these cultures produces friction without leverage.
The deliverables are not residency-shaped. Pure infrastructure engineering work (cost optimization, latency tuning, retrieval refactoring) often does not benefit from co-location; the work is the same in both regimes. Residencies pay back when the work has a strong domain-context dependency; they do not pay back when the work is largely independent of domain.
For engagements that fall into any of these patterns, the right shape is the standard remote engagement with a sharper kickoff and tighter cadence; not a forced residency.
FAQ
The rotating residency model is not a marketing offering; it is a delivery mechanism that, when fit, produces materially better outcomes than the standard remote engagement. The key is fit: engagements with strong domain dependency, expansion ahead, and the budget to amortize the residency overhead are the ones where the model pays back. Engagements that run residencies for the optics produce no leverage; engagements that run them for the structural advantage produce the most successful Q3 quarters in the cadence.
Arthur Wandzel is the founder of SFAI Labs, a forward-deployed AI development agency in San Francisco. He has run roughly a dozen residencies across portfolio companies and clients in the last 18 months.
Frequently Asked Questions
What is an AI agency residency?
A 4-to-6-week period during which a single named senior agency engineer works on-site (or in a structured remote-residency arrangement) embedded inside the client engineering organization. The engineer attends client standups, reviews client PRs, pairs with client engineers on eval design, and is co-located with the client domain expert for the duration. Their work artifacts ship into the client repo, not into the agency’s repo and then transferred. The residency has a fixed end date and bounded deliverables; typically one feature shipped, one eval expansion completed, one knowledge-transfer artifact produced.
Why does a residency outperform a documentation handoff?
Four reasons. First, fastest knowledge transfer; documentation captures explicit knowledge well and tacit knowledge poorly, and a residency engineer absorbs the unwritten rules in a week. Second, deepest data understanding; AI systems are dominated by data understanding, and watching the client team handle real cases produces a different mental model than reading a doc. Third, real-time eval pairing; eval design is a high-bandwidth collaboration that benefits from same-room presence. Fourth, builds relationships; the residency engineer becomes the person the client calls during stress events in month seven.
When in the engagement should a residency happen?
Most powerful in Q3 of the quarterly mandate cadence. By Q3 the system is in production, the eval suite is mature, and the open question is expansion; adding a second feature, integrating a second data source, extending to a second user segment. Q3 is when the agency engineer’s deep client-context understanding has the highest leverage, and the residency builds the relationship surface that supports a Q4 transition. Q1 residencies are sometimes appropriate but more often premature; Q2 residencies are rarely the highest-leverage placement.
Who from the agency should rotate into a residency?
The two roles that most benefit from a residency are the lead engineer and the eval owner. The lead engineer rotation places the senior engineer who owns the architecture in the client building for the architectural decisions of the quarter. The eval owner rotation places the engineer who maintains the eval suite next to the domain expert who provides ground-truth cases. The lead engineer rotation is per-quarter, which spreads client-context understanding across the agency rather than concentrating it in one person. The eval owner rotation is per-domain when an engagement spans multiple domains.
How is a residency different from a body-shopping arrangement?
Five structural boundaries. The agency remains employer of record (the residency engineer is paid by the agency, has agency benefits, and reports to an agency manager). The residency has a fixed end date in writing. The deliverables are bounded and named; typically three named artifacts. The residency engineer does not become part of the client on-call rotation. The residency does not displace client hiring; it is a knowledge-transfer mechanism, not a substitute for the client building their own AI engineering capability.
How is a residency priced?
As a fixed-fee deliverable inside the relevant quarterly mandate, not as time-and-materials. The fee covers the engineer’s time at the agency rate (typically 1.2 to 1.4x the engineer’s standard hourly rate to account for residency overhead), plus named travel and accommodation costs at receipts, plus the residency artifacts delivered. Fixed-fee residency pricing aligns incentives; a time-and-materials residency creates an incentive for the residency to extend, while a fixed-fee residency creates an incentive for the engineer to deliver and exit on schedule. The residency premium over a non-residency week is typically 15-25%.
Can a residency be done remotely?
Yes, with conditions. Remote residencies work when three conditions hold: the residency engineer’s working hours overlap with the client’s by at least 6 hours, the client has a synchronous-communication culture with strong video presence in standups and willingness to pair on calls, and the deliverables can be unambiguously produced without same-room presence. Remote residencies typically have 80 to 90 percent of the leverage of on-site at 60 to 70 percent of the cost. Hybrid residencies; two weeks on-site, two weeks remote; work well when the client is geographically distant.
What does a healthy on-site cadence look like during a residency?
Typically 4 days per week in the client office, 1 day remote for agency-internal work. A full 5 days on-site loses the agency-side context the engineer needs to maintain; they need synchronous time with their agency lead, time to write up artifacts in their own working environment, and a structural break from client immersion. The 4-and-1 rhythm preserves the residency’s leverage while keeping the engineer connected to the agency’s broader context, code reviews, and team conversations.
When is a residency the wrong fit?
Three patterns. First, the engagement is too small; a residency on a 90-day engagement with a single feature in scope is overkill, and residencies typically fit engagements with $200K+ in quarterly fees. Second, the client is hostile to embedded engineering; some cultures prefer asynchronous collaboration, have security models that resist on-site contractors, or have executive sponsorship uncomfortable with the optics. Third, the deliverables are not residency-shaped; pure infrastructure engineering work like cost optimization or latency tuning often does not benefit from co-location because the work is the same in both regimes.
What happens if the residency engineer leaves the agency mid-engagement?
The rotation structure handles this. Each quarter is a different lead engineer, so client-context knowledge is distributed across multiple agency engineers rather than concentrated in one. The artifacts the residency produces; eval expansions, architecture decisions, knowledge-transfer documents; live in the client repo and persist independent of any single engineer. The relationship surface is broader than any individual residency. If a residency engineer departs mid-residency, the agency rotates a replacement with a 5-business-day handoff window and absorbs the cost of the transition rather than passing it to the client.
Arthur Wandzel
