⟶ Insights / AI Engineering
Custom AI agents that survive production.
How senior teams design custom LLM pipelines, tool-using agents and RAG systems that hold up under real load, real compliance and a real invoice - not just a demo reel.
5 min read · By the Jogiitech engineering team · Updated July 2026
Design premise
The failure mode of custom agents is not intelligence - it is unbounded action. Every production agent we ship is scoped by tool allow-lists, step budgets and typed output contracts before a prompt is written.
The winning architecture is boring on purpose: retrieval that is measured, tools that are versioned, prompts that are diffed, evals that gate every deploy. Cleverness lives in the workflow, not in the model call.
Frequently asked.
What is a 'custom AI agent' in production terms?+
An LLM-driven workflow that can read context, call tools (databases, APIs, internal services), plan multi-step actions, and produce structured output that a downstream system can consume. In production it also ships with evaluation, guardrails, a reviewer UI and a rollback path. Anything without those four is a prototype.
RAG, fine-tuning, or agents - how do we choose?+
RAG when the answer lives in your corpus and freshness matters. Fine-tuning when tone, format or a narrow judgement task is repeatable and evaluable. Agents when the workflow spans multiple systems, needs planning, and a single prompt cannot do it. Most real systems blend all three: RAG for retrieval, fine-tuned smaller models for classification steps, and an agent loop for orchestration.
How do we stop agents from hallucinating or going off-task?+
Five layers. (1) Strict output schemas - validate every step against a typed contract. (2) Tool allow-lists - the agent can only call the tools it has been granted. (3) Step budgets and cost caps - hard limits per run. (4) Eval harness with regression tests on every prompt or model change. (5) A human-review UI on high-stakes actions until confidence is measured, not assumed.
What about prompt injection and data exfiltration?+
We treat any retrieved or user-supplied text as untrusted. Defences include content isolation (retrieved data is never mixed into the system prompt), egress allow-lists on tools, PII redaction on both inbound and outbound payloads, and dedicated red-team eval sets that live in CI. Prompt injection is a design-time problem, not a runtime patch.
How do we control the token bill?+
A model gateway sits between the app and the providers, so we can route by task class - frontier model for reasoning, smaller open-weight for extraction, cache for repeats. We instrument cost per request, per user and per workflow from day one, and set alerts on unit-economics regressions before they hit the invoice.
How long until a custom agent is in production?+
A scoped single-domain agent (one workflow, one team, real users) ships in 10-14 weeks with a supervised rollout. Multi-agent systems with reviewer UI, escalation flows and audit trails typically take 16-24 weeks. Discovery workshop first, always - we do not quote agent work without a two-week scoping engagement.
Build versus buy, honestly.
Not every workflow needs a bespoke agent. The decision comes down to how proprietary the workflow is, how much it costs to run at your volume, and how much control you need over data residency and behaviour. Use this matrix as a starting point, then stress-test it against your actual traffic and compliance constraints.
| Factor | Favours buying a platform | Favours a custom-built agent |
|---|---|---|
| Workflow uniqueness | Generic - transcription, generic chat, generic OCR that many vendors already do well. | Proprietary to your business - the workflow itself is a source of competitive advantage. |
| Data sensitivity | Standard data with an acceptable vendor DPA and no residency constraint. | Regulated or highly sensitive data that needs on-prem or VPC-isolated processing. |
| Volume and unit economics | Low to moderate volume where per-seat or per-call vendor pricing is still cheaper than engineering time. | High volume where a custom pipeline with model routing and caching meaningfully undercuts vendor pricing. |
| Time to first value | Days to weeks - you need something running now and can iterate on process around it. | Weeks to months is acceptable because the payoff compounds over a multi-year roadmap. |
| Integration depth | The workflow sits at the edge of your systems and a SaaS connector covers it. | The workflow needs deep, two-way integration with internal systems, tools and approval chains. |
| Long-term ownership | You are comfortable being dependent on a vendor's roadmap and pricing changes. | You need to own the logic, the data and the ability to change behaviour without a vendor's release cycle. |
Most enterprise agents we ship sit on the build side of this table for the core workflow, while buying commodity components - transcription, embeddings, base models - around it. Our custom AI engineering practice runs this matrix explicitly in the discovery workshop, so the build decision is on the record before any code is written.
Permissions, audit trails and human approval.
The moment an agent can write - update a record, issue a refund, send an email, trigger a payment - the engineering bar changes. Read-only agents fail gracefully; a bad answer is embarrassing. Write-capable agents fail expensively; a bad action is a support ticket, a compliance incident, or worse.
We treat write access as a permission to be granted narrowly and logged completely, not a capability that follows from the agent being 'smart enough'. In practice that means: every write action maps to a specific, scoped credential rather than a shared service account; every tool call is logged with the input, the output and the identity of the agent run that made it; and any action above a defined risk threshold - amount, reversibility, regulatory exposure - routes to a human approval step before it executes, not after.
The audit trail is not an afterthought bolted onto logging - it is a first-class design requirement. A reviewer should be able to answer 'why did the system do this' from the trail alone, without pulling in an engineer to explain the prompt. That means storing the retrieved context, the model's stated reasoning where available, the tool calls made, and the final action, all tied to a single traceable run ID. For anything touching money, personal data or contractual terms, we default to human-in-the-loop until the agent's approval-override rate on a stratified sample has stayed low for a defined observation period - confidence is earned from evidence, not assumed from a demo.
Scoping a custom agent?
30 minutes with a senior architect. We will map the workflow, the guardrails and the first eval set before you sign anything.