Project Management

Please login or join to subscribe to this thread

Spec-driven development with AI agents: is Agile still working for your team?

linkedin twitter facebook   Agile   Artificial Intelligence   Scrum  
avatar
Anonymous

A team adopting spec-driven development (SDD) with AI coding agents. Coding is much faster, but writing and reviewing specs is now the bottleneck, and fixed sprints, story points, and standard ceremonies feel out of sync.

For those managing similar projects:

  1. Did you keep Scrum, move to Kanban/Scrumban, or change something else? What exactly did you adjust (sprint length, ceremonies, estimation, roles)?
  2. How do you plan, forecast, and report now? Which metrics do you track?
  3. How do you manage spec and review capacity, and stakeholder expectations?
  4. Which tools are you using (Jira, Linear, Spec Kit, Kiro, MCP/agents), and what worked or didn’t?

Real examples and “it didn’t work for us” stories are very welcome. Thanks!
Sort By:
avatar
Gulam ishakkhan G OPERATIONS ENGINEER| HINE HYDRALICS INDIA PVT LTD Villupuram, TN, India
Disclosure up front: I’m sharing how I’d approach this from a PM and flow perspective [add: and what I’ve seen on my own projects, if any].
1. Scrum vs. Kanban/Scrumban
When the constraint moves from coding to specifying and reviewing, I’d lean toward Scrumban:
  • Drop story-point estimation. Instead, right-size specs so each is small enough to review in one sitting.
  • Replace the fixed sprint commitment with WIP limits on the stages that matter: spec drafting, spec in review, and agent implementation.
  • Keep a regular cadence (1 to 2 weeks) for stakeholder demos and retros, but make planning continuous rather than a single event.
  • Add explicit roles or responsibilities: who owns the spec, who reviews it, and who signs off on acceptance criteria.
2. Planning, forecasting, reporting
Story points and velocity lose meaning when implementation time collapses. Flow metrics fit better:
  • Lead time and cycle time per spec, from idea to merged
  • Throughput (specs completed per week), used for probabilistic forecasting such as Monte Carlo
  • Review queue age and time-in-review, the real bottleneck signal
  • Spec rework rate (how often a spec is revised after agent output)
  • Defect escape rate, to confirm speed isn’t hiding quality problems
3. Managing spec and review capacity
  • Treat review capacity as the scarce resource and plan around it, as you would a constrained supplier.
  • Tier reviews by risk: lightweight review for low-risk changes, deeper review for architecture, security, and data changes.
  • Use spec templates with required acceptance criteria so reviews check against a standard.
  • Set review SLAs and protect focus time for reviewers.
  • With stakeholders, report outcomes and delivered value, not “how fast the AI codes.” Speed in coding doesn’t mean speed in decisions.
avatar
Lissette Indhira Pimentel Sosa
Community Champion
Program Manager| HARPER SRL Santo Domingo / Distrito Nacional, Dominican Republic
I think this is where we need to reconsider what we're actually measuring. If AI agents accelerate coding but specifications and reviews become the bottleneck, tracking story points or development velocity alone may give us a misleading picture.
I'd pay closer attention to end-to-end cycle time, review turnaround, work in progress, and rework rates.
The goal is to adapt our planning and metrics to where the work is actually slowing down.

Please login or join to reply

Content ID:

"I don't know much about being a millionaire, but I'll bet I'd be darling at it."

- Dorothy Parker

Sponsors

Newsletters

Nov 4: Fix It Yourself? Maybe You Shouldn’t...
Oct 28: Leadership Isn't a Popularity Contest
Oct 21: Leadership Takes Courage

See all newsletters