Novo Agents vs building it yourself
The DIY stack is genuinely good now: model APIs, agent frameworks, durable-execution infrastructure, and sandbox services are all excellent. The catch is that none of them is the agent. This page costs the assembly honestly, and says when building is the right call.
What you sign up to build
Wiring a model API into your product gets you text back. Between that and an agent your users can trust sits the part nobody's roadmap mentions:
- an orchestration harness that plans, delegates, and retries
- sandboxed browser and workspace infrastructure
- document pipelines for Office files, PDF, and export
- web research with untrusted-content fencing
- durable execution: crashes, deploys, and reconnects
- pause-and-resume for human approval
- per-run metering, budgets, and receipts per customer
- re-tuning all of it every time the models change
What the evidence says
- Quality is the production blocker. In LangChain's State of Agent Engineering survey (1,340 respondents, fielded late 2025), quality was the most-cited barrier to shipping agents, ahead of latency and cost.
- The harness moves the score more than you think. LangChain raised its own agent 13.7 points on Terminal-Bench 2.0 with the model held fixed, purely by harness engineering (February 2026). Academic work in 2026 found harness-induced variance can exceed model-induced variance, including cases where model rankings reverse between harnesses.
- Containment is a real liability. An independent assessment of 100 production agents, analyzed in a Cloud Security Alliance research note (June 2026), found 98% carried simultaneous private-data access, untrusted-content exposure, and outbound actions, and only 11% passed a baseline security benchmark.
- Cost attribution stays unsolved. Only 22% of finance leaders can tie AI spend to business outcomes (CloudZero/Sapio survey, February 2026, n=500). If you resell your product, per-customer agent cost is margin survival, and you would be building that ledger too.
The part that never finishes
The build is not a project with an end date. Harness technique moves quarterly, models deprecate on published clocks measured in months, and every model change re-prices your prompts, your tool designs, and your eval suite. The teams that win at DIY treat the harness as a permanent engineering competence. That is exactly what Novo sells: the harness, the sandboxes, and the model routing are our whole product, so they never decay into your roadmap.
What you keep either way
Buyers consistently refuse to outsource prompts, domain tools, evals, and UX, and they are right. Novo's split matches that line: your identity and context, your domain tools over MCP or OpenAPI, your transcript, your UI. Novo runs the loop, the managed capabilities, the durability, and the billing. Tools execute in your infrastructure through signed adapters, and provider keys never enter it.
When building it yourself wins
- The harness is your product. If agent quality is the differentiator you sell, own it. A skilled team with eval discipline can beat any platform's ceiling for its domain.
- You need total architectural control. No vendor boundary, custom replay semantics, exotic compliance surfaces.
- Unit economics at hyperscale. Past a certain volume, owning the stack at raw compute rates can price below any managed layer, if the engineering payroll is already sunk.
await novo.threads.stream(thread.id, { input: task }).Statistics above are primary-sourced and dated 2026-07-15 in our internal research; sources: LangChain State of Agent Engineering and harness-engineering blog, the CSA research note on the AIRQ agent assessment, and CloudZero's finance survey.