The framework
Every component a production agent platform needs.
BCG’s enterprise agent research names ten components a platform has to provide before agents can run in production. Here is what each is for, and who delivers it when you run Katonic.
of technology leaders fear silent failure on AI investments: spending without real impact.
BCG, 2025The ten at a glance
- 9 delivered by Katonic
- 1 your data platform
The ten components are BCG’s. Sorting them into build, run and govern is Katonic’s. Select a tile to jump to it.
● In productionePLDT, the ICT arm of the Philippines’ largest telco, chose Katonic to power Pilipinas AI, the nation’s first sovereign AI solutions stack, with a full Workroom, Studio and Control Room deployment.Read the case →
Why a framework
The model is not the hard part. Everything around it is.
After two years of building agents with clients, BCG’s finding is that what holds agents back is not the language model but the legacy systems and processes around it. It lists five blockers. Its answer is a shared platform: build each capability once and let every team reuse it.
- 01
Legacy integration
Agents have to be stitched into old stacks, mixed APIs and fine-grained access rules, and each join brings security and change-control risk.
- 02
Unreliable data
Siloed, low-trust, slow-moving data makes an agent’s decisions brittle. Scale needs data that is clean, current and well governed.
- 03
Missing evaluation
Long reasoning paths hide where an agent went wrong. Tracing tool calls, red teaming and testing on broad data is real work.
Answered by LLMOps, Monitoring
- 04
Governance and audit
Enterprises need explainability, guardrails and policy compliance from the first day, which adds work up front.
Answered by Guardrails, Monitoring
- 05
Operating at scale
Leaving the pilot stage needs ownership, incident handling, control of cost and latency, versioning and change tracking.
The five blockers are BCG’s, paraphrased. Which components answer each one is Katonic’s reading.
The ten components
What each one is for.
And who delivers it.
In italics: what BCG says the component is for, in our words.In plain type: what Katonic ships today.
- 01
AI Guardrails
KATONICGuardrails offered as a shared service to every generative AI application, so no team has to build its own.
Three layers: content safety, trajectory instrumentation and a governance proxy with per-organisation policy and human approvals. PII scanning, jailbreak detection, grounding checks and topic filters come built in.
See Guardrails →Back to the ten ↑ - 02
LLMOps
KATONICManages the prompt lifecycle, evaluates agents, and gives observability across everything an agent did on the way to its answer.
Prompts are versioned in Studio. Built-in evaluators, versioned datasets and promotion gates block a release when scores drop, and every model call is traced with its token count and cost.
See Evaluation →See Studio →Back to the ten ↑ - 03
MCP and Agent Registry
KATONICOne place to store MCP server and tool definitions and agents, and to make them available to everyone who builds.
240+ governed MCP tools, agent versioning, and publish, promote and rollback.
See MCP servers →Back to the ten ↑ - 04
Model Gateway
KATONICA single access layer for model endpoints, so the rest of the platform never deals with quotas or scaling. Models can be switched without changing code, and spend is tracked at one point of control.
Any model your IT approves, behind one OpenAI-compatible endpoint, with bring-your-own-key per team, budget enforcement and health-aware failover.
See AI Gateway →Back to the ten ↑ - 05
No/Low-code Builders
KATONICLets people create and customise agents quickly through a visual interface and drag-and-drop parts.
Five build paths: a guided builder, a conversational builder, a workflow canvas, an IDE with the SDK, and bring your own agent. Every path ends on the same governed runtime.
See Build an agent →Back to the ten ↑ - 06
Agent Framework
KATONICCoordinates multi-agent workflows: planning, selecting the right agent or tool, and routing by policy.
LangChain, CrewAI, AutoGen, LlamaIndex or the native runtime. Teams already shipping agents on LangChain, CrewAI or LangGraph can bring them in.
See Agent architecture →Back to the ten ↑ - 07
Agent Runtime
KATONICRuns the agents, holds session state, and enforces policy while they run.
Production execution, hot-reload and immutable agent versions. Three isolated environments, with promotion from dev to test to prod behind evaluation gates and one-click rollback.
See Environments →Back to the ten ↑ - 08
Memory
KATONICLets an agent recall past actions and behaviour: short-term memory keeps one session coherent, long-term memory carries knowledge across sessions.
Short-term context, long-term semantic search, and isolation per organisation.
Back to the ten ↑ - 09
Monitoring, Logging, FinOps
KATONICContinuous traceability, cost visibility and evaluation, so prompts, outputs and spend can all be managed.
Per-agent observability, cost attribution and a full audit trail. Cost is tracked per agent, per department and per model.
See Control Room →Back to the ten ↑ - 10
Data Platform
YOUR DATAThe structured and unstructured data sources agents are allowed to reach. BCG treats it as a capability that stays critical once agents arrive.
Yours. Your existing Snowflake, Databricks, SAP, SharePoint or S3. We connect through 50+ enterprise connectors with permission-aware retrieval, and never duplicate.
See Connectors →Back to the ten ↑
Also in BCG’s diagram
BCG’s picture of the platform draws a few more boxes around the ten. Here is where Katonic stands on each.
- Vector and knowledge stores
- Knowledge Engine: hybrid search across vector, keyword and graph, with retrieval that respects each user’s permissions. Knowledge Engine
- Traditional machine learning models
- Studio: build, fine-tune and ship models alongside your agents. Studio
- Identity and access
- Four role tiers, single sign-on, human approvals and a full audit trail across tenants, teams and tools. Governance
- Secure code sandbox and browser access
- Not counted in the nine. If your agents need to run code or drive a browser, raise it on the call.
Choosing a platform
How BCG says to choose. Where Katonic fits.
Buy the foundations
BCG’s default is an off-the-shelf platform, with custom building kept for the use cases that set you apart. Its first step is to assess the stack you already have and extend it, not replace it.
Nine of the ten come out of the box, connected to the data platform you already run.
Expect a hybrid
BCG sees no one-size-fits-all platform. Most enterprises will buy agents embedded in their main software suites and build the orchestration that cuts across them.
Agents built in other frameworks can call Katonic’s knowledge, published agents and governed tools over MCP, under the same governance as internal calls.
Follow the data
Of the forces BCG weighs, data access pulls hardest: agents should sit where enterprise data lives, because moving it adds latency, fragility and cost. Existing systems come second, governance and compliance third.
Katonic runs on infrastructure you control: your own data centre, your own cloud account, or fully offline.
Next step
Score your stack
against the ten.
Bring your current stack to a thirty-minute call. We go through the ten components with you and show the nine Katonic delivers running on infrastructure you control.
Source
Framework, component definitions and platform guidance adapted and paraphrased from BCG, “Building Effective Enterprise Agents” (PDF on bcg.com), AI Platforms Group, November 2025. The 75% figure is from the same briefing, which cites BCG market research. The mapping to Katonic, the grouping into build, run and govern, and the count of nine and one are Katonic’s own. BCG has not reviewed or endorsed this page.