AI-Powered Developer Productivity Platform with MCP Servers and Agent-Based Automation

bloomberg · Finance · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.0 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise developer productivity platform

Bloomberg needed to help more than 9,000 engineers adopt AI coding, code-uplift, and incident-response agents without duplicating tooling, creating chaotic MCP servers, or losing governance over models and internal systems.

Constraints:

Bloomberg has one of the world's largest JavaScript codebases
Many teams want similar bots and incident-response agents
AI-generated patches still require deterministic verification and human review

Architecture

Bloomberg’s developer-productivity effort is a platform/paved-path architecture for enterprise AI adoption. Teams use an AI gateway to compare and select models with central visibility, an MCP directory/hub to discover and partner on existing MCP servers, and a standard PaaS for MCP tool creation, SDLC, runtime, and auth. On top of that platform, uplift agents scan codebases and propose patches, while incident-response agents reach metrics, logs, topology, alarms, triggers, and SLOs through MCP servers. Adoption is reinforced through onboarding, AI productivity communities, champ/guild programs, and leadership workshops.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  engineering-teams[Engineering teams] --> ai-gateway[AI gateway]
  engineering-teams --> mcp-directory-hub[MCP directory hub]
  mcp-directory-hub --> mcp-server-platform[MCP server platform]
  mcp-server-platform --> mcp-servers[MCP servers]
  mcp-servers --> observability-systems[Observability systems]
  ai-gateway --> code-uplift-agents[Code uplift agents]
  code-uplift-agents --> codebase-and-pr-workflow[Codebase and PR workflow]
  ai-gateway --> incident-response-agents[Incident response agents]
  mcp-servers --> incident-response-agents
  onboarding-training[Onboarding training] --> engineering-teams
  ai-productivity-community[AI productivity community] --> engineering-teams
  leadership-workshops[Leadership workshops] --> engineering-teams

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`ai-gateway`	Central gateway for model experimentation, model-usage visibility, and platform guidance on model fit.	`engineering-teams`, `code-uplift-agents`, `incident-response-agents`
`mcp-directory-hub`	Discovery layer where teams find existing or in-progress MCP servers and partner instead of duplicating work.	`engineering-teams`, `mcp-server-platform`
`mcp-server-platform`	PaaS for MCP tool creation/deployment, SDLC support, runtime provisioning, authentication, and authorization.	`mcp-directory-hub`, `mcp-servers`
`mcp-servers`	Connectors to internal systems such as metrics/log dashboards, topology, alarms, triggers, and SLOs.	`mcp-server-platform`, `incident-response-agents`, `observability-systems`
`observability-systems`	Metrics, logs, dashboards, network/service topology, alarms, triggers, and SLO sources used in incident workflows.	`mcp-servers`
`code-uplift-agents`	Agents that scan codebases, identify applicable patches/refactors, and create pull requests with rationale.	`ai-gateway`, `codebase-and-pr-workflow`
`codebase-and-pr-workflow`	Large software asset base and PR review/merge process where generated patches land and create downstream review load.	`code-uplift-agents`
`incident-response-agents`	Agents that gather code, telemetry, feature flags, traces, topology, alarms, and SLO data to support troubleshooting.	`ai-gateway`, `mcp-servers`
`onboarding-training`	Formal training path that teaches new engineers AI coding practices with Bloomberg principles and technologies.	`engineering-teams`
`ai-productivity-community`	Cross-organization champ/guild community for shared learning, deduplication, inner-source, and visiting-engineer collaboration.	`engineering-teams`, `mcp-directory-hub`
`leadership-workshops`	Enablement program for technical leaders whose AI-tool adoption lagged individual contributors.	`engineering-teams`
`engineering-teams`	The 9,000+ engineer population that consumes the paved path and contributes use cases/tools.	`ai-gateway`, `mcp-directory-hub`, `onboarding-training`, `ai-productivity-community`, `leadership-workshops`

Data flow

A team with an AI coding or operations use case starts on the paved path. It can experiment with models through the AI gateway, look in the MCP hub before building a new server, and deploy MCP tooling through the standard platform service with runtime and auth handled centrally. Uplift agents then scan codebases and generate patch PRs; incident-response agents traverse code, telemetry, feature flags, traces, logs, topology, alarms, and SLOs through MCP servers. Training and communities feed adoption patterns back into the organization so duplicated work is reduced and useful tools spread.

4 supporting quotes

“we have a gateway so that teams can easily figure out which model works the best. They can do quick experiments.”transcript, stated

“we have a two discovery uh basically MCP directory via hub so that let's say team A wants to do something they will go to the hub”transcript, stated

“tool creation and deployment is via a pass it's basically a um you know a standard platform service where you can do your STLC and and we provide runtime environment for you”transcript, stated

“the key part is we're going to need to build a lot of MCP servers to connect to the um the metrics and logs dashboards”transcript, stated

Why this architecture

Bloomberg chose a paved-path platform because a 9,000-engineer organization would otherwise create duplicate PR bots, incident-response agents, and MCP servers. The gateway and hub preserve experimentation while making model usage/tooling visible; the MCP PaaS centralizes runtime and auth; and tiered demo-vs-production controls keep ideation easy while requiring quality controls for production in a reliability-sensitive business.

4 supporting quotes

“you have 10 teams who wants to build a pull request review bots. you have too many teams who wants to build a instant response agents right they become very quickly chaotic and sometimes can have duplications”transcript, stated

“we kind of believe in provide a golden path um with enablement teams”transcript, stated

“we want to make demo very easy so that or I really say prove concept very easy so that people can try have ideal generation”transcript, stated

“we also want to make sure the production requires some quality control. um because at the end of the day stability and system reliability is at the core of our business.”transcript, stated

Models used: LLMs

Scale

The platform serves an engineering organization of more than 9,000 engineers, mostly software engineers.
“We have more than 9,000 engineers and most of them are software engineers.”transcript, stated
Bloomberg’s software estate includes one of the world’s largest JavaScript codebases, making uncontrolled AI patching risky.
“We also have one of the largest JavaScript codebase um in the world.”transcript, stated
The incident-response use case spans many operational systems, so MCP servers are a platform concern rather than one-off connectors.
“connect to the um the metrics and logs dashboards you have connect to the topology you have whether it's network topology or it's the um your service dependency topology uh your alarms your triggers right your SLOs's”transcript, stated
The adoption program uses long-running institutional channels: a 20-year training program and decade-old champ/guild communities.
“in Bloomberg we have a wellestablished training program uh it's more than 20 years”transcript, stated
“within Bloomer we have something called um a champ program and a guild program”transcript, stated

Alternatives considered (roads not taken)

Uncoordinated team-by-team agent and MCP server creation

Rejected because: At Bloomberg scale, many teams wanted the same classes of bots and servers, producing duplication and chaos.

“too many teams who wants to build a instant response agents right they become very quickly chaotic and sometimes can have duplications”transcript, stated

A maximally prescriptive central platform where every business unit must use the same abstraction

Rejected because: Bloomberg positions itself between team autonomy and tight abstraction, preferring a golden path with enablement teams.

“Some company allow teams to be have a lot of freedom”transcript, implied

“some organization have a very very strong tight abstraction of the service infrastructure”transcript, implied

Regex-based refactoring/uplift tools only

Rejected because: They worked only to a limited extent, while LLM/tooling-based uplift produced better results for semantic code evolution.

“We did have a reg based refraction tool um it works to some extent but it's limited right now with um LMS and other tooling. So we are able to uh see very much better results”transcript, stated

Treating POCs and production access the same way

Rejected because: Creativity needs easy demos, while production requires quality controls because reliability is core to the business.

“creativity come from some freedom of try different new things but we also want to make sure the production requires some quality control.”transcript, stated

Patterns

`tool-use-orchestration` (enterprise-mcp-platform)

Bloomberg standardizes MCP server discovery, creation, deployment, runtime, auth, and connection to metrics/logs/topology/SLO tools.

“we're going to need to build a lot of MCP servers to connect to the um the metrics and logs dashboards”transcript, stated

`model-routing` (ai-gateway)

The AI gateway lets teams experiment with models, gives platform visibility into model usage, and guides teams toward fit-for-purpose models.

“we have a gateway so that teams can easily figure out which model works the best. They can do quick experiments.”transcript, stated

`retrieval-ranking` (tool-discovery)

The MCP directory/hub helps teams discover existing servers and partner instead of duplicating tools.

“we have a two discovery uh basically MCP directory via hub so that let's say team A wants to do something they will go to the hub”transcript, stated

`canary-rollout` (enterprise-adoption)

Bloomberg inserts AI coding into onboarding and uses communities/guilds to spread practices organically.

“we just incorporate AI coding in on boarding training program and also show them how to best utilize them with our principles and our technologies”transcript, stated

`eval-driven-development` (code-uplift-quality)

The code-uplift lesson is that deterministic tests, linters, and verification are essential before generated patches can be trusted.

“it would be really nice if there's some detistic verification capability. uh oftentimes it's not so easy especially if you have test cases if you don't have good llinter”transcript, stated

`cost-aware-routing` (engineering-economics)

Bloomberg treats AI tooling as changing which engineering work is cheap or expensive enough to pursue.

“it actually changes the cost function of software engineering.”transcript, stated

Tradeoffs

How much platform freedom to allow.

Chose a golden path with enablement teams over letting every business unit build arbitrary AI/MCP infrastructure.

Rationale: The platform makes the right thing easy and the wrong thing difficult while still allowing team-level use cases.

“we kind of believe in provide a golden path um with enablement teams”transcript, stated

“we want to make sure the wrong thing is ridiculous hard to do.”transcript, stated

How to coordinate shared tooling.

Chose MCP hub and inner-source/guild collaboration over parallel duplicate implementations by separate teams.

Rationale: Discovery and community programs let teams partner, share learning, and deduplicate effort.

“organically it dduplicates efforts and there's shared learning uh shared learning happening”transcript, stated

How to separate experimentation from production use.

Chose easy demo/POC access plus production quality controls over the same access standard for all AI tools and MCP servers.

Rationale: Bloomberg wants creativity and ideation in early use, but production systems need controls because stability and reliability are central to the business.

“we want to make demo very easy so that or I really say prove concept very easy so that people can try have ideal generation”transcript, stated

“we also want to make sure the production requires some quality control.”transcript, stated

How to spread AI practices across a large engineering organization.

Chose onboarding, AI productivity communities, champ/guild programs, and leadership workshops over tool deployment without organizational enablement.

Rationale: The speaker describes training and communities as mechanisms that create comfort, deduplicate effort, and compensate for leadership adoption gaps.

“we just incorporate AI coding in on boarding training program and also show them how to best utilize them with our principles and our technologies”transcript, stated

“organically it dduplicates efforts and there's shared learning uh shared learning happening”transcript, stated

Gotchas (what bit them)

Teams building similar agents can quickly create duplication and chaos.

Root cause: Many teams independently want PR review bots, incident-response agents, and MCP servers.

Resolution: Bloomberg created a PaaS/golden path with gateway, MCP hub, and standard deployment services.

“you have 10 teams who wants to build a pull request review bots. you have too many teams who wants to build a instant response agents”transcript, stated

Code uplift agents can increase open PR volume and time to merge.

Root cause: Generated patches still need review and merging, so the review bottleneck can move downstream.

Resolution: Treat deterministic verification and review capacity as part of the rollout, not an afterthought.

“the average open pull requests increased and time to merge also increased uh because you spin a lot of new code and then still we have to review the code”transcript, stated

Leadership adoption lagged individual contributors.

Root cause: ICs adopted coding tools more strongly than leaders, creating a need for targeted change management.

Resolution: Bloomberg used onboarding, communities, champ/guild programs, and shared learning to drive adoption.

“our data shows individual contributors have a much better stronger adoption than our leadership team.”transcript, stated

Evaluation

Bloomberg tracks adoption, model/tool usage through the gateway, duplicated effort, onboarding effectiveness, open PR/time-to-merge effects, and qualitative changes in engineering tradeoffs.

Metrics: AI coding adoption · model usage visibility · duplicated effort · open pull requests · time to merge · community participation

“there's a lot more f familiarity and comfort with the tooling. Um and also the important part is there's a lot more nuance insights of where it's at value”transcript, implied

“we can have visibility what kind of models being used and we can also guide through teams which model should is a better fit”transcript, stated

Outcomes

Bloomberg created an AI gateway, MCP directory/hub, and standard MCP deployment service.
AI coding was incorporated into onboarding training.
AI productivity communities helped deduplicate effort and share learning.
The platform reframed engineering tradeoffs by changing the cost function of software engineering.

Bloomberg has more than 9,000 engineers, one of the world's largest JavaScript codebases, and many teams building AI tooling across terminal, infrastructure, and reliability workflows.

“We have more than 9,000 engineers and most of them are software engineers.”transcript, stated

“We also have one of the largest JavaScript codebase um in the world.”transcript, stated