From SDLC to agent development lifecycle

How must your enterprise architecture think to support them? Because the central truth of the new era is simple: You cannot deploy cognitive systems on non-cognitive infrastructure.

Why do CIOs need to make a transition from systems development to agent development?

For the last two decades, enterprise technology has functioned on one straightforward presumption. Applications follow workflows. Workflows follow requirements. Requirements follow the business.

This is the logic that shaped traditional SDLC: Systems Development Life Cycle. A linear chain of planning, design, implementation, testing, deployment, and maintenance. It worked because systems were deterministic, data models were static, and human operators filled every cognitive gap. In an agent mediated environment, however, this logic no longer holds.

Dynamic workflows are evolving, real-time data sources are shifting, customer behaviors are changing at an hourly basis, and the systems have to make probabilistic decisions out at the edge of incomplete information. SDLC cannot support this. ADLC can.

My experiences in finance, retail, supply chain, HR, and customer ops shows that when enterprises attempt to deploy agents on the architecture of the SDLC era, there are three recurring failure patterns

• Agents are flexible, while workflows are brittle. SDLC workflows presuppose single correct path; agents consider multiple paths and select based on outcome and constraint. This yields "exception overload" as old systems attempt to force deterministic logic on probabilistic cognition.
• Integrations expect well structured data, whereas agents need multimodal grounding. OCR, screenshot, chat log, supplier PDFs, customer emails, invoice images… These inputs invalidate the basic assumptions of ETL. Semantic basis is necessary for agents in all formats, including text, graphics, signals, metadata, and policies at the same time.
• SDLC assumes predictable requirements, while agent systems evolve continuously. An agent that handles returns today may, through data, grounding and evaluation loops, become capable of fraud detection next month. Under SDLC, this creates backlog, governance issues, and architecture debt. Under ADLC, it becomes a feature of continuous improvement.

What ADLC actually is and why it matters now?

Across Salesforce's Agent Development Lifecycle, Microsoft's reimagined developer lifecycle and Sierra's operational agent loops, a common pattern emerges. ADLC is not a replacement for SDLC, it is a cognitive overlay and it runs continuously, not sequentially.

Enterprise grade agents use 5 live loops.

1. Planning: basically, intent interpretation and task decomposition. The agent converts intents like broad "process this claim", "prepare a response", "analyze these logs" into atomic tasks. This shifts business rules from BPM diagrams into the internal reasoning engine of the agent.
2. Knowledge then Context then Constraints. Grounding transforms the enterprise environment into a navigable map; structured data, semi-structured supplier files, PDF contracts, call logs, operational KPIs, governance constraints, customer history, external signals. This layer determines whether the agent "understands" the real state of the enterprise, and it is where most organizations discover their data and integration debt.
3. Actions are performed through API calls, triggering of workflows, RPA, or system to system orchestration. The agent selects from probable actions and determines risks while staying within the boundaries set on the allowed actions. In enterprise reality, that is where failsafes, rate limits, audit logs, and guardrails come in.
4. After every action, the agent evaluates whether the goal of the action was achieved. Other types of evaluation include validation against business rules, verification of compliance, nomaly detection, human escalation triggers, and general task progress. The evaluation loop is where agent reliability is built.
5. Agents would evolve on the fly using user corrections, regret signals, performance telemetry, growing knowledge bases, domain fine tuning, and semantic memory updates. Under SDLC, this would be "scope creep." Under ADLC, it's become an intelligence.

Insights: Where is the CIO struggling most?

Based on my industry discussions with leaders, some consistent friction points emerge, and from client observations, I would rather not mention OCR and extraction pipeline issues anymore.

(a) Leaders recognize that agents should be "modular," but the majority of enterprises are still operating monolithic logic, legacy schemas, and tightly coupled decision engines. Plus, agent modularity requires data, actions, and policies to be fully modular.

(b) Data backbones cannot support multimodal cognition. Structured + semi-structured + unstructured rarely live in a unified substrate. This is the No #1 barrier to enterprise grade agents. This is the largest blind spot CIOs face and the defining requirement for cognitive systems.

(c) Distributed workloads hit the interconnectivity and latency ceilings. Agents need to retrieve context rapidly and reason across services fast. Distributed AI workloads face bottlenecks that no previous legacy SDLC had ever needed to consider; agents also require near real-time memory access, policy constraints, and multi-service reasoning things. Legacy integration stacks cannot keep pace with.

The following actions break the SDLC limitation

First, rather than designing for applications, design for: What does the agent perceive? What does the agent remember? What is the agent allowed to act on? Where does the agent require human intervention? This framing often exposes latent data gaps and coordination inefficiencies. Second, construct the agent trust layer (KYA) In my previous post https://www.garudax.id/pulse/agent-cognition-architecture-cios-ibrahim-ot-05nwf I have attempted to present the trust architecture; here I tried to how to make it more operational. Authentication, authorization, policy constraints, semantic logging, decision traceability. All these might be very crucial at the beginning of the journey in order to get maximum trust from your outcomes against your client safe action boundaries. Finally, create a unified multi-modal grounding substrate. Next post will go deep into this, we will talk about semantic knowledge graphs, redesigning existing pipelines for multimodal inputs, high-throughput interconnects, context windows on top of enterprise memory.

If you are a CIO or CDO, the next 12–18 months demand a shift from;

• systems design to cognitive design
• integration pipelines to grounding substrates
• process governance to agent governance
• change requests to learning loops

Because the enterprise is no longer a set of workflows. It’s an emergent cognitive environment. Next post will complete the picture with answering the question:

Why multimodal grounding is the first true competitive advantage in the agent era and what it takes to build it?

If you're leading technology in finance, logistics, retail, or a telco, and you're already mapping where agentic AI could unlock operational speed or cost advantages, feel free to reach out. I've built small, sector-specific opportunity maps that outline the 3-6 highest-ROI agent use-cases for each industry.

Just a practical overview you can use to pressure test your roadmap in less than 10 minutes.

Resources

• Salesforce, The Agent Development Lifecycle: From Conception to Production 2025
• Microsoft, Agentic DevOps in action: Reimagining every phase of the developer lifecycle, 2025
• The Agent Development Life Cycle — Zack Reneau-Wedeen, Sierra

#agenticAI #DigitalTransformation