What Is Scrum

February 6, 2025 · Baoss
BAOSS IT consulting: Scrum with AI agents, from ceremonies to value

Scrum in 2025: From Ceremonies to Real Value with AI and Agents

In 2019, when we published the first version of this article, the question was “What is Scrum?”. In 2025, the question CTOs and VPs of Engineering ask us at BAOSS is radically different: “We have Scrum, Jira, ‘agile’ teams and certifications… why are we still delivering late, with technical debt and no predictability?” From BAOSS IT consulting, the answer is no longer more ceremonies. Instead, it is AI agents that provide context and evidence.

The short answer: Zombie Scrum. That is, teams that run the ceremonies (Daily, Planning, Review, Retro) but have disconnected the framework from business outcomes. According to the State of Agile Report 2025 (Digital.ai), 68% of organizations admit their teams “do Scrum” without achieving real agility. In addition, the Standish Group’s 2025 CHAOS report updates its historical metric: “agile” projects only triple the success rate versus waterfall when they combine technical excellence, measured value flow and cognitive automation via AI. Without that triangle, Scrum is just expensive theater.

The real 2025-2026 problem BAOSS IT consulting diagnoses: cognitive load, invisible dependencies and “Feature Factories”

Today, an average Product Owner manages backlogs of 300+ items. A Scrum Master spends 60% of their time chasing Jira updates instead of removing systemic impediments. Also, developers lose 15-20 hours/sprint to context switching between code, PRs, refinement and cross-team alignment meetings. That’s why the complexity is not in the framework (which fits in 13 pages in the Scrum Guide 2020), but in the sociotechnical environment: coupled monolithic architecture, data fragmented across Confluence/Jira/GitHub/Slack and an absence of automated feedback loops.

  • Blind planning: Sprint Plannings based on gut feeling and historical velocity, ignoring technical risk, cross-dependencies and real capacity (holidays, on-call, debt).
  • Inefficient refinement: Hours writing generic acceptance criteria (ACs), poorly sliced stories (vertical slicing nonexistent) and lack of coordination with UX/Architecture.
  • Opacity within the Sprint: The Daily becomes a status report. Impediments are detected late. The Sprint Goal breaks mid-iteration with no early warning.
  • Performative review: Demo-centric Sprint Reviews, with no adoption metrics, business value (Outcome over Output) or hypothesis validation.

At BAOSS we have audited 40+ transformations in 2024-2025 (banking, logistics, healthtech, retail). The pattern is universal: **Scrum exposes the problems but doesn’t solve them**. Therefore, the 2025-2026 change lever is injecting **agentic artificial intelligence** into the Sprint lifecycle to automate the mechanical, augment the cognitive and measure what matters. Disruption gives no warning: is your company exposed to the impact of AI agents?.

BAOSS Case #1 (anonymized): B2B Fintech – From 6h Planning to 90 min with RAG + Planning Agents

Context: 8 teams (50 devs), migration from monolith to microservices. 6-8 hour Sprint Planning, a Say/Do ratio (committed vs delivered) of 52%, high PO-Dev friction due to ambiguous stories.

Intervention: We deployed a RAG system (Retrieval-Augmented Generation) over their Jira history (3 years), Confluence (ADRs, specs), GitHub (PRs, commits) and Slack (technical channels). On top of that vectorized knowledge base (embeddings with text-embedding-3-large + pgvector), we orchestrated a multi-agent planning agent (LangGraph) with three specialized nodes. Also, with agents reading past decisions, AI-powered cybersecurity stops being optional: the next attack on your company will no longer be human.

  • Backlog Analyst (Claude 4 Opus): Reads the Sprint Backlog candidates, cross-references them with ADRs and known technical debt, proposes technical-functional vertical slicing and generates executable Gherkin-style ACs.
  • Capacity Simulator (GPT-4o + Code Interpreter): Ingests calendars (Outlook), on-call rotations (PagerDuty), historical lead time by task type and component. Output: probabilistic range of real capacity (P10/P50/P90) per team and skill.
  • Dependency Validator (AutoGen Agent + MCP Jira/GitHub): Walks the graph of Issue Links, open Pull Requests and DB schema changes. Alert: “Story H-4421 blocks Teams 3 and 5 unless PR-882 is merged before Wednesday”.

Results after 6 sprints (3 months):

  • Planning time: 6h → 90 min (async agent preparation + 30 sync minutes of alignment).
  • Say/Do Ratio: 52% → 87% (commitment calibrated by the simulator).
  • Story Quality (Defect Leakage): –38% production bugs from incomplete ACs (measured in Sentry/Jira).
  • PO/Dev Satisfaction (internal eNPS): +42 pts.

“The agent doesn’t decide for us; it brings us augmented reality. We discuss *risks*, not *tasks*.” — Lead Product Owner, B2B Fintech

BAOSS Case #2 (anonymized): Logistics/Retail – Eliminating cross-team blockers with Autonomous Agents (CrewAI + MCP)

Context: 12 teams in Scaled Scrum (Nexus/lightweight LeSS). 40% of Sprint Goals failed due to dependencies not detected until the Daily or the Review. Average Lead Time: 22 days. Architecture: modulith transitioning to cell-based architecture.

Intervention: We implemented a **Dependency Resolution Swarm (CrewAI)** connected via **MCP (Model Context Protocol)** to Jira (Portfolio/Advanced Roadmaps), GitHub (CODEOWNERS, branch protection), ArgoCD (deployment state) and Datadog (SLIs/SLOs). That is, the crew is composed of:

  • Scout (explorer agent): Scans the active Sprint Backlogs every 15 min, detecting overlaps of modified files/DBs/APIs (dynamic vs static code ownership).
  • Negotiator (negotiator agent): Proposes integration sequence diagrams (Mermaid.js) and suggests contract tests (Pact) or feature flags (LaunchDarkly) to decouple deployments.
  • Committer (executor agent – human in the loop): Creates technical sub-tasks, links issues, opens draft PRs with contract test templates and notifies CODEOWNERS on Slack.

Results after 4 months:

  • Cross-team blockers/sprint: –63% (from 18 to 6.6 on average).
  • Deployment Frequency: 1.2/week/team → 3.8/week/team (DORA Elite).
  • Lead Time for Changes: 22 days → 9 days (P50).
  • MTTR (Mean Time To Recovery): 4.2h → 47 min (thanks to auto-suggested feature flags).

Technical key: **MCP standardizes context**. The agent doesn’t “guess” the state of the main branch or the ArgoCD configuration. On the contrary, the MCP server serves it to the agent typed. This enables reliable reasoning (CoT – Chain of Thought) in Claude 4 Sonnet / GPT-4o without hallucinations about infrastructure. Also, every agent you connect to Jira or GitHub expands the risk of AI agents: define boundaries and control before deploying.

The new mental map: “Augmented” Roles, Events and Artifacts

The Scrum Guide doesn’t change, but the professional implementation does. Here is how we operate in 2025-2026:

Roles → Accountabilities + AI Copilots

  • Product Owner → Product Leader: From

    Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.