Micro Agency vs Mid-Size Agency: How Big Should Your Software Partner Be in 2026?

TL;DR

For a typical custom build, the team touching your code should be about three to seven people, and the firm should be large enough to name a second person for every critical part of the system. QSM's analysis of 491 completed projects found three to seven person teams performed best, with effort growth turning sharply non-linear from around nine people, and Brooks's n(n-1)/2 explains why. AI changes how much a small team produces, not the coordination maths: Google's 2025 DORA research found AI adoption correlated positively with throughput and negatively with delivery stability, and a Q1 2026 survey of 250 agencies found founder posture, not headcount, predicted whether agents reached production. The one-person agency fails on calibration and continuity rather than output, since a solo engagement guarantees a truck factor of one. Four contract clauses close the gap: a named second person per critical component, disclosed subcontractors with flow-down obligations, named domain reviewers, and repository, cloud and licences held in your own accounts.

There is a loud argument running through agency land right now. One camp says the one-person agency is the future, because a single operator with autonomous agents covers strategy, design, code and QA. The other camp says the real winners are micro agencies, three to five experienced people using AI to deliver what a forty person shop used to, and that mid-size agencies with departments, processes and handoff rules are the ones in trouble.

Both camps are arguing about their own business model. If you are the one buying software, the question is different and much more useful: how many people does your partner actually need, and what do you lose at each size? This article answers that with the delivery-research evidence, the 2026 agency economics, and the contract clauses that decide whether a small team is a bargain or a single point of failure.

Comparing a small studio against a bigger shop?

Get a Second Opinion

How big should a software development partner be?

For a typical custom build, the team touching your codebase should be about three to seven people, and the firm should be large enough to name a second person who can carry every critical part of the system. Below that, you are buying one person's availability. Above it, extra headcount mostly buys peak capacity, specialist bench and procurement comfort rather than delivery speed. AI changes how much a small team can produce. It does not change the coordination maths, and it does not remove the need for a second experienced human to disagree with the first.

What does the delivery evidence say about team size?

The strongest number in this debate predates the AI cycle by more than a decade. Quantitative Software Management analysed 491 completed projects in the 35,000 to 95,000 new-or-modified source-lines-of-code band, average project size 57,412 effective SLOC, and grouped them into five team-size buckets.

Team size	What the 491-project sample showed
1.5 to 3	Low total effort, but weaker schedule performance than the middle buckets
3 to 5	Best productivity index, close second on schedule
5 to 7	Best schedule performance, productivity a close second
7 to 9	Measurably lower productivity, higher effort
9 to 11	Effort growth turns sharply non-linear

QSM's summary is blunt: a three to seven person team had the best performance for medium-sized information systems, the smaller teams scored two or more productivity indices above the larger ones, and the extreme non-linear effort increase does not kick in until team size approaches nine or more. Separate QSM work comparing best-in-class against worst-in-class projects found the top performers used teams roughly four times smaller on average.

The mechanism is old and unglamorous. Brooks described it in 1975: pairwise communication paths grow as n(n-1)/2, while software tasks rarely partition cleanly. That formula is worth internalising before you read a proposal that promises twelve people.

People on the delivery team	Pairwise communication paths
3	3
5	10
7	21
12	66
40	780

A forty person shop does not run 780 conversations. It buys process, departments and handoffs to suppress most of them, and that suppression is exactly what buyers experience as slow change control, account managers who did not write the code, and a two week wait for an answer a developer could have given in ten minutes.

Why is AI compressing agency headcount in 2026?

Because AI removes a large share of the routine execution that justified junior layers, and it removes it faster than agencies can restructure around it. Forrester's 2026 predictions put a number on the adjustment: after an average 8% agency headcount cut in 2025, it forecasts a 15% reduction in agency jobs in 2026, and expects agencies to change business models rather than merely shrink.

The margin data points the same way. Promethean Research surveyed digital agency owners and managers in February 2026 (N=119) for its 2026 State of Digital Services report. Average agency size in the sample was 31 employees, and average after-tax net margin was 13% in 2025. Split by size, the pattern is the interesting part.

Agency size	Average after-tax net margin, 2025
Studios under 10 employees	19%
Industry average	13%
Agencies with 50 or more employees	8%

Promethean also counts more than 71,000 digital agencies in North America, of which 87% employ fewer than 50 people. So the mid-size squeeze thesis has real numbers behind it, and small is already the norm rather than the exception.

Read those numbers honestly. Forrester's forecast and Promethean's survey both cover marketing and digital agencies, not custom software engineering firms, and the Promethean sample is 119 self-selected respondents. They tell you which way the economics lean. They are not a benchmark for a software delivery partner, and anyone quoting them at you as one is overclaiming.

Does agency size predict how well a partner uses AI?

No, and this is the finding buyers should act on. A Q1 2026 survey of 250 agencies with 8 to 180 full-time staff found 41% had at least one agent in production, up from 9% a year earlier, with a median self-reported ROI of 3.2x. Its own conclusion about what separated the leaders was that the cluster was determined by founder posture and willingness to rebuild the delivery process, not by revenue or headcount. Two forty person agencies in the same sample had completely different agent deployments.

The same survey is a useful lesson in reading vendor AI claims. When the authors validated self-reported figures against billing data from 47 agencies, ROI was over-reported by roughly 18% and token spend under-reported by roughly 24%, and they noted the sample over-indexes on agencies already interested in agentic AI. Treat any "we deliver 10x with AI" pitch as a claim that needs evidence, whether it comes from a solo operator or a mid-size shop. Our own field notes on running coding agents inside a token budget describe what real leverage looks like in practice: cache design, routing and context discipline, not a headline multiple.

Google's 2025 DORA research, based on nearly 5,000 technology professionals, frames the ceiling on all of this. Around 90% reported using AI at work and more than 80% said it improved their productivity, yet about 30% reported little or no trust in AI-generated code. DORA found AI adoption correlated positively with software delivery throughput and negatively with delivery stability. Its headline conclusion is that AI is an amplifier: it magnifies whatever the organisation already is. A disciplined micro team amplifies discipline. A partner with weak review habits amplifies those instead, faster.

Where does the one-person agency actually break?

Not on output. On calibration and continuity.

Calibration. METR's controlled early-2025 study of experienced open-source developers working in repositories they knew well found they took 19% longer with AI assistance, while forecasting a 24% speedup beforehand and reporting a perceived 20% speedup afterwards. Whatever the current size of the real effect, and METR's 2026 update explains why that has become hard to measure cleanly, the perception gap is the durable finding. AI raises confidence faster than it raises correctness. The cheapest available correction is a second experienced person willing to say "this design will not survive the second customer." That is a structural argument for two or more senior humans, and it is the part of the original micro agency argument that holds up best.

Continuity. The research term is truck factor, or bus factor: the number of people who would have to disappear before a project stalls. Avelino and colleagues estimated it across 133 popular GitHub applications and found 65% had a truck factor of two or fewer, with 46% sitting at one. Their estimates matched developer reality in 84% of validation responses. A one-person agency does not merely risk a truck factor of one on your project. It guarantees one, contractually, for as long as the engagement runs.

That is survivable for some work. A four week integration, a prototype, a well-scoped internal tool: a strong solo operator is often the fastest and cheapest route, and pretending otherwise would be dishonest. It is not survivable for a system that has to run in production for years, be handed to someone else, or pass technical due diligence. The gap between those two cases is the whole buying decision.

Solo vs micro agency vs mid-size vs large: a buyer's comparison

Partner shape	What you actually get	Where it breaks	Fits best
Solo operator 1 person	Direct access to the person doing the work, no coordination cost, lowest rate for equivalent seniority	Truck factor of one, no peer review, no cover for illness or a competing client, no specialist depth	Prototypes, single-service integrations, short well-defined scopes, augmenting a team that already has senior review
Micro agency 3 to 7 people	Senior peer review inside the team, the QSM productivity sweet spot, the person who scoped the work is the person who builds it	Limited surge capacity, specialist gaps covered by subcontractors, procurement questionnaires and certification demands are hard to satisfy	Custom products, MVPs that must reach production, multi-year systems with a small dedicated team, rescue and takeover work
Mid-size agency 20 to 80 people	Multiple parallel workstreams, formal roles, established process documentation	Coordination overhead, seniority dilution on your account, the pitch team is rarely the delivery team, thinnest margins in the 2026 data	Programmes needing several parallel teams, buyers whose procurement requires organisational depth
Large firm 200+ people	Staffing guarantees, certifications, on-call rotas, contractual muscle, multi-country rollouts	Cost, change latency, and a real chance your work is delivered by the most junior available bench	Regulated multi-year programmes, tenders with formal vendor requirements, anything needing 20 or more engineers within weeks

If you have not yet settled the prior question of sourcing model, our freelancer vs agency vs in-house guide covers that decision with Austrian cost scenarios. This article assumes you have chosen to buy from an outside firm, and are now deciding how big that firm should be.

The four ways a small partner fails, and the clause that covers each

Every objection to hiring a three to seven person team reduces to one of four risks. Each has a specific, checkable contractual answer. If a small partner cannot give you these four in writing, the objection is valid. If they can, most of the size argument evaporates.

Risk	What you ask for	What a real answer looks like
Key person unavailable	A named second person per critical component, and a documented current state	Two names against each of database, deployment, authentication and integrations; architecture decision records in the repository; a written response time for when the primary is unreachable
Peak capacity	Disclosed subcontractors under the partner's own statement of work, with flow-down obligations	Named specialists with role, location and access level; confidentiality, security, IP and GDPR Article 28 processor terms flowed down; one contract and one invoice; the partner still accountable for quality
Specialist depth	Who reviews security, privacy and scaling, by name and by evidence	A named reviewer per domain and the artifact they produce, plus an honest statement of which certifications the firm does not hold
The firm disappears	Ownership and operability independent of the partner	Source control, cloud and CI in your organisation's accounts, licences in your name, credentials in your vault, and a tested handover package rather than a promise of documentation

The fourth row is the one buyers systematically skip and later pay for. Our software handover checklist is the artifact-level version, and the software agency proposal teardown shows how these obligations get quietly negotiated away in a proposal that otherwise reads well. If you are already mid-engagement and the answers are missing, the 30-day provider change plan is the recovery path.

How do you vet a micro agency in one call?

Ten questions, in this order. The pattern to listen for is whether an answer names a person, an artifact or a date. Answers made of adjectives are not answers.

Who writes the code on my project, and are they on this call? In a genuine micro agency the answer is yes. In a mid-size shop it usually is not.
Who reviews their work, and what happens when the two of you disagree about a design? The absence of a named reviewer is the single biggest gap in small-team delivery.
Show me your last handover. Redacted is fine. Vagueness here predicts lock-in.
Where do AI agents run in your delivery, and what does a human check before merge? Look for review gates, tests and a policy on generated code, not tool names.
What is your test and release process for my kind of system? Compare the answer against a QA checklist before launch rather than against their confidence.
Which parts of this build are outside your depth, and who covers them? A firm that claims no gaps has one it has not told you about.
What happens to my project if your primary engineer is out for three weeks? Ask for the mechanism, not the reassurance.
Whose accounts hold the repository, cloud, CI and domain? The answer should be yours, from week one.
What would make you decline this project? A partner with no declining criteria is selling capacity.
How is change priced and approved? Fixed price and time and materials both work. Unowned change control does not.

The fuller vetting sequence, including reference calls and how to shortlist, lives in our guide to choosing a software agency.

When is a bigger partner still the right answer?

Small is not automatically better, and the honest version of the micro agency argument has limits. Choose scale when:

you need more than roughly twenty engineers working in parallel within weeks, and the work genuinely partitions;
procurement requires certifications, audited processes or insurance floors a small firm does not hold;
you need a contractual round-the-clock on-call rota rather than best-effort senior availability;
the programme spans several countries, regulators or business units at the same time;
your own organisation will not assign a decision-maker, in which case you are buying process and account management, and you should buy them deliberately.

That last one is worth saying plainly. A small senior team is fast because decisions happen in one conversation. If nobody on your side can make those decisions, the speed advantage disappears and the coordination layer you skipped has to be rebuilt on your side of the contract.

How Wavect is structured, and why

Wavect is a founder-led studio, deliberately. Both managing directors ship, so the people who scope the work are the people who build it, and design disagreements happen between two senior humans before they reach your codebase. Specialists we have shipped production work with before join under Wavect's own statement of work as subcontractors, which means one contract, one invoice and one accountable party rather than a referral into a network. Repository, cloud and CI live in the client's accounts from the start.

We do not hold ISO 27001 and we do not offer a contractual round-the-clock SLA. If your procurement process requires either, a larger firm is the correct choice, and we will say so on the first call. When it does not, the trade you are making is fewer people against senior continuity and direct access. See how that plays out on a real system in the Bond Analytics case study, or compare it side by side against a generalist development agency and against hiring through freelance platforms.

Frequently Asked Questions

What is a micro agency?

A micro agency is a firm of roughly three to seven experienced practitioners with no separate delivery layer: the people who sell the work also do it. It differs from a solo operator because peer review, cover and a second senior opinion exist inside the team, and from a mid-size agency because there are no departments, account managers or internal handoffs between the client and the builder.

How many developers should a software project have?

For a medium-sized system, three to seven people on the delivery team is the evidence-backed range. QSM's analysis of 491 completed projects in the 35,000 to 95,000 source-lines-of-code band found three to seven person teams performed best, with three to five strongest on productivity and five to seven strongest on schedule, and the extreme non-linear effort increase only kicking in from around nine people.

Can a one-person agency deliver production software?

For narrow, short scopes, often yes. The structural problems are calibration and continuity: no peer review of design decisions, and a truck factor of one for the whole engagement. Research on 133 popular open-source applications found 65% already had a truck factor of two or fewer, so this is a common failure mode rather than a hypothetical one. For a system that must run for years or pass due diligence, insist on a named second person.

Is a small software agency riskier than a big one?

It carries different risk. A small firm concentrates key-person and surge risk; a large firm concentrates cost, change latency and the risk that your work lands on its most junior available bench. Both are manageable in contract. For a small partner, cover four specifics: a named second person per critical component, disclosed subcontractors with flow-down obligations, named domain reviewers, and ownership of repository, cloud and licences on your side.

Does AI mean a smaller agency can deliver what a big one used to?

Partly, and not because of headcount. A Q1 2026 survey of 250 agencies found that what separated agencies with agents in production was founder posture and willingness to rebuild the delivery process, not revenue or headcount. Google's 2025 DORA research found AI adoption correlated positively with delivery throughput but negatively with delivery stability, and concluded that AI amplifies existing organisational strengths and weaknesses. Small plus disciplined scales well. Small plus undisciplined fails faster.

Why are mid-sized agencies under the most pressure?

They carry the coordination cost of scale without the contractual advantages of scale. Promethean Research's February 2026 survey of 119 agency leaders found studios under ten employees averaged 19% after-tax net margin in 2025 against 8% for agencies with fifty or more, and Forrester forecasts a 15% cut in agency jobs in 2026 after an average 8% cut in 2025. Both datasets cover marketing and digital agencies rather than software engineering firms, so read them as direction, not as a software benchmark.

What is bus factor or truck factor, and why should a buyer care?

It is the number of people whose sudden unavailability would stall a project. It matters because it is the one vendor risk that contract language can fix cheaply and that buyers almost always discover too late. Ask for two names against each critical component, decision records in the repository, and a tested handover package rather than a promise of documentation.

How do I compare a micro agency quote against a mid-size agency quote?

Normalise both to first-year comparable cost and to who does the work. Add excluded work, hosting, licences, your own team's required effort and expected exit cost. Then compare the seniority of the people actually assigned, the review arrangements, named subcontractors and the handover obligation. A lower headline price from a larger firm often reflects a more junior assigned team.

Sources and how they were checked

Every figure in this article was verified at the primary source on 18 August 2026. The Forrester and Promethean datasets describe marketing and digital agencies rather than software engineering firms, so they are used as direction of travel and not as a software delivery benchmark.

QSM, Team Size Can Be the Key to a Successful Software Project: 491 completed projects, 35,000 to 95,000 SLOC, five buckets from 1.5 to 11 people, best performance at three to seven people.
QSM, Top Performing Projects Use Small Teams: best-in-class projects used teams roughly four times smaller on average.
Forrester, Predictions 2026: Marketing Agencies Resign Their Agency: after an average 8% headcount cut in 2025, a forecast 15% reduction in 2026.
Promethean Research, 2026 State of Digital Services: surveyed February 2026, N=119, average agency size 31 employees, 13% average after-tax net margin in 2025.
Google DORA, State of AI-assisted Software Development 2025: nearly 5,000 technology professionals, AI as an amplifier, positive relationship with throughput and negative relationship with delivery stability.
METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity: 19% longer despite a forecast and perceived speedup, with the February 2026 update on why the effect is now hard to measure cleanly.
Avelino, Passos, Hora and Valente, A Novel Approach for Estimating Truck Factors: 133 popular GitHub applications, 65% with a truck factor of two or fewer and 46% at one.
Digital Applied, Agentic AI Adoption Survey Q1 2026: 250 agencies of 8 to 180 staff, 41% with at least one agent in production, self-reported figures validated against billing data from 47 agencies.

Final thoughts

The one-person agency versus micro agency argument is a debate between sellers. The buyer's version is simpler, and older than the AI cycle: three to seven people is where delivery research says the work goes best, and the threshold that matters is whether a second experienced person can carry each critical part of your system.

AI moves how much output a small team can produce. It does not move the coordination maths, and the evidence suggests it widens the gap between what people believe they shipped and what they actually shipped. That is an argument for senior peer review, not for removing the second human. Size your partner for review and continuity, put the four risks in writing, and the headcount question mostly answers itself.