It is one of the most frequent decisions in technical teams starting with AI, and one of the most often taken with an incomplete calculation. The argument for building is a good one: the team has capable people, the frameworks are mature and open, and a platform license is a recurring cost. The argument is not wrong. It is incomplete, and the missing part is the part that decides.
What "building" actually includes
When somebody says they are going to build an agent on a framework, what they have in mind is the agent: the instructions, the tools, the model answering. That is the visible part. For that agent to run in production with volume and risk, the team also builds:
- Observability: a record of every execution, every decision and every tool call with its parameters and return, reproducibly. Application logging does not do the job.
- Access control and human approval: who may trigger which action, under which limit, and where a person approves before the agent acts.
- Model governance: routing between providers, fallback when one goes down, a cost limit per team, and what never leaves the company before reaching the model.
- Memory and context management: what persists, for how long, how it is summarized so it does not overflow, and how it is erased when the data subject asks.
- Sustainable integration: connectors with per-user authentication, error handling, idempotent retries. Every system is a project.
- Versioning and promotion: how a change to the instructions goes from test to production without breaking what is live, and how you roll it back.
- Operation: a dashboard, alerts, sample review, continuous tuning. Somebody has to watch.
None of those seven items is the agent. All of them are mandatory for the agent to exist in production. And the framework, by definition, hands you the parts to build them, not them ready. A "build" calculation that only includes the agent is counting a seventh of the work.
When building is the right choice
Worth being clear, because it happens: building from scratch is the correct decision in some situations, and in those a platform would be cost with no return.
- The case is one of a kind and will not produce a second. One agent, one process, no plan to expand.
- The team has a senior engineer available, with time, and AI is their priority, not the third task of the week.
- The process is documented and the data is accessible. The only uncertainty is technical.
- There is no deadline pressure. The time until the first agent is in production costs the operation nothing.
- The requirement is so specific that no platform covers it, and adapting one would cost more than building.
When all five are true at once, build. The comparison with LangChain and LangGraph details which scenario the framework is the answer to and which it is not.
What gets left out of the cost comparison
The usual comparison puts a platform license on one side and two engineers' salaries on the other, and stops there. What is missing from the build column:
The time until production
While the team builds the operations layer, the operation keeps paying for the old process. That cost does not appear on the project spreadsheet, but it is real and it is monthly. Four months of difference in go-live are four months of the spending the agent exists to reduce.
The second agent
The first agent justifies assembling everything from scratch. The second reveals that half of what was assembled was generic and should have been reused, but was not designed for it. The fifth reveals that the team accidentally built a platform, with no documentation, no roadmap and a dependency on whoever wrote it.
The person who leaves
A framework is chosen by one person and configured their way. When they leave, the knowledge goes with them, and the next engineer inherits a system only the author understood. That risk has no line on the spreadsheet and it is what most often turns a project that worked into one nobody wants to touch.
The first audit
Security, privacy or the enterprise customer will ask for a reproducible trail, access control by action and evidence that sensitive data does not leave untreated. Building that afterwards, under audit pressure, costs many times what building it alongside would have.
The third option the question hides
The question "build or buy" assumes that buying means receiving a tool and fending for yourself. There is a third option, and it is the one that most often reaches production: a platform plus engineers who build with your team, inside your operation, until it runs.
In that model the operations layer comes ready, the specific case is built by people who know the platform alongside people who know the process, and your team is trained during the deployment instead of receiving a handoff at the end. What you buy is not software: it is arrival in production, with the knowledge staying in house.
A practical rule when evaluating any option: ask who sustains the agent six months after go-live, and what happens when that person is not around. The answer says more about the real cost than any proposal.
How to decide
| If your scenario is | The choice tends to be |
|---|---|
| One case of its own, a senior team available, no deadline, a documented process | Build on a framework. |
| Several cases foreseen, a technical team busy with the core business | A platform. |
| A first project, an undocumented process, a deadline that matters | A platform with engineers deploying alongside. |
| A regulated sector, with audit requirements from the start | A platform with native governance, whatever the deployment model. |
| A requirement no platform covers | Build, and plan for the second agent from day one. |
Where Runflow comes in
Runflow is the third option: the platform to build, operate and measure agents, plus the Forward Deployed engineers who deploy with your team until it runs in production, and who stay after go-live. The seven items on the list above come ready in the platform, and the TypeScript SDK exists so that your team builds in code what is specific, without rebuilding what is common. When your scenario is the one for building on a framework, the comparison says so, and so do we.
