Executive Intelligence · Build vs Buy: AI Agents | October 4, 2026
Good morning, Folks! Once the model decision is made, the next one sits at the application layer: who builds the agents that actually do the work? It is not simply build or buy. There is a third path: configure, and for many enterprises it is the right place to start. This episode of our Build vs Buy series maps all three.
A demo agent is a model plus a prompt. A production agent is a system, and the model is often the smallest part of it.
Integration, permissions, and reliability are where most agent projects succeed or stall. That is why the build-vs-buy question here is really a question about where your engineering effort should go.
What an Enterprise AI Agent Actually Requires
Models — One or more that reason and plan.
Tools and integrations — Access to CRMs, ERPs, ticketing, email, and databases.
Memory and knowledge — Usually retrieval over company data.
Orchestration — Logic to sequence steps, handle errors, and retry.
Permissions and identity — The agent can only do what its human principal could do.
Guardrails, logging, and human approval — Observable, auditable behavior, with sign-off on consequential actions.
The CODEW Lens: The model is the brain. The plumbing, permissions, and oversight are the body, and that is where the budget goes.
Three Paths: Build, Buy, or Configure
Build — Your engineers assemble agents on LLM APIs or open-weight models, using orchestration frameworks and your own integration code. Maximum control, maximum responsibility.
Buy — You adopt agents embedded in software you already use (CRM, service desk, HR, finance) or from specialized vendors. Fastest to deploy, least differentiating.
Configure — You use an agent platform that supplies orchestration, connectors, and governance, and you define the workflows. A middle path trading some flexibility for speed and built-in controls.
The CODEW Lens: Configure is the path most debates forget, and often the one that fits.
Horizontal vs. Vertical Agents
Horizontal agents (general assistants, research, coding, meeting and email agents) serve many functions and are usually best bought. Vendor scale advantages are large, and your use is rarely unique.
Vertical agents (claims, underwriting support, clinical documentation, procurement, contract review) embed domain workflows and compliance logic. A specialist may beat a custom build if they have already solved your industry's edge cases. Building wins when the workflow itself is your competitive advantage.
The CODEW Lens: Buy what everyone does. Build what only you do.
The Build Case
Building fits when the process is proprietary and encoding it in a vendor's product risks commoditizing it; when legacy systems or strict permission structures defeat off-the-shelf connectors; when you need exact control over approval, logging, and audit; and when high-volume repetitive work makes owning the stack cheaper than per-seat or per-action pricing. It also lets you route tasks across model providers as they improve.
When Building Wins — Three Conditions:
1. Workflow is the differentiator. The process itself is proprietary.
2. Integration is unusual. Off-the-shelf connectors cannot reach your systems safely.
3. You can maintain it. Engineering capacity exists to keep up with model and tooling changes.
The CODEW Lens: Build when the agent encodes something that makes your business distinct.
The Buy and Configure Case
Buying or configuring wins on time to deployment (weeks, not quarters, with visible results to fund further investment), offloaded maintenance of frameworks and connectors, built-in governance such as audit trails and role-based access, and embedded context: an agent inside your CRM already has the data model, permissions, and interface. It also lowers the dependence on scarce AI engineers, though you still need people who understand the workflows.
The CODEW Lens: A vendor's agent starts with your data model already in hand. A custom one has to earn it.
Security, Governance, and Reliability
Agents differ from chatbots because they take actions, which changes the risk profile.
Least privilege — Narrow, task-specific permissions instead of broad service accounts.
Prompt injection — Agents that read emails, web pages, or documents can be manipulated into unintended actions. Test for it deliberately.
Approval gates — Human sign-off for irreversible or high-value actions: payments, deletions, external communications.
Observability — Log every tool call and decision so incidents can be reconstructed.
Evaluation — Maintain realistic task suites and rerun them whenever the model, prompt, or tools change. Silent regressions are common.
The CODEW Lens: With vendors, ask what is logged, what you can configure, and what happens to your workflows if the product changes or disappears.
Cost, ROI, and Time-to-Deployment
| Factor | Build | Configure | Buy |
|---|---|---|---|
| Time to value | Slowest | Moderate | Fastest |
| Upfront cost | Highest | Moderate | Lowest |
| Ongoing cost | Engineering and maintenance | Platform fees plus workflow upkeep | Licenses, per seat or usage |
| Customization | Highest | Moderate | Lowest |
| Lock-in | Your own stack | The platform | The vendor's product and data model |
Measure ROI on outcomes, not activity: cycle time reduced, cases resolved without escalation, error rates, hours returned to higher-value work. Build the measurement before deployment, or value cannot be proven afterward.
The CODEW Lens: If you cannot measure the "before," you will never be able to claim the "after."
Multi-Agent Architectures and Lock-In
Systems where specialized agents collaborate (planner, researcher, executor, reviewer) can handle complex workflows, but they multiply failure modes, cost, and debugging difficulty. Start with one well-scoped agent, prove it, and add agents only when a clear limitation demands it.
Own your assets — Keep business logic, prompts, and evaluation sets in formats you control.
Stay multi-model — Prefer platforms that support several model providers.
Negotiate the exit — Secure data export and exit terms before signing and favor open protocols for tool connections where available.
The CODEW Lens: More agents is not more intelligence. It is more surface area.
Build vs Buy: The CODEW Verdict
Six Questions — In Order:
1. Is this workflow a source of competitive advantage? If yes, lean build. If commodity, buy.
2. Does an embedded agent already exist in software you use? Try it first.
3. How complex are the integrations and permissions? High complexity favors build or configure.
4. What are the risk and compliance stakes? High stakes demand control you can verify.
5. Do you have the capacity to maintain agents through model and tooling changes?
6. How fast do you need results? Under a quarter means buy or configure first.
The CODEW Lens: Buy for commodity work, configure for common workflows with some customization, and build only where the agent makes your business distinct.
The Build vs Buy Glossary
Agent — A system that uses a model to plan and take actions through tools, not just generate text.
Human-in-the-Loop — A control requiring human approval before consequential actions.
Orchestration — The logic that sequences steps, tools, and retries.
Prompt Injection — Malicious instructions hidden in content an agent reads, intended to redirect its behavior.
Least Privilege — Granting an agent only the permissions its task requires.
Vertical Agent — An agent built around one industry's workflows and compliance needs.
FAQ
Q: Should we build agents on our own LLM infrastructure?
Usually the model layer is bought, and the agent layer is where you choose. Build the agent only if the workflow is distinctive and you can maintain it.
Q: What is the difference between buying and configuring?
Buying adopts a finished agent. Configuring uses a platform's orchestration, connectors, and governance while you define the workflow logic.
Q: Are multi-agent systems worth it?
Only when a single agent demonstrably hits a limit. Each added agent multiplies cost and failure modes.
Q: What is the biggest risk with agents?
Unintended action: excess permissions combined with manipulated inputs. Narrow access, approval gates, logging, and continuous evaluation are the defenses.
The CODEW Stat
3 paths · 6 questions · 1 rule Build, buy, or configure. Six questions sort the decision. One rule closes it: build only where the agent encodes what makes your business distinct.
Reviewed by Erwin Castro
on
Sunday, October 04, 2026
Rating:
No comments: