Why Enterprise Agents Can’t Close Jira Tickets

▶ Watch (05:33)

McKinsey put a number on the problem: 88% of companies use AI, but only 6% see value creation. Navakoti’s explanation is the three-tier Jira ticket. Green tasks use general LLM knowledge. Orange tasks need procedural guidance you can teach. Red tasks depend on tribal knowledge inside people’s heads, never written down. Agents handle green and orange. They stall on red.

Enterprise teams respond by building 10 to 20 MCP servers against a knowledge base where 40% is tribal, 20% is outdated, and 20% is unreliable. More servers don’t fix that. They just surface the same gaps faster.

Pull Instead of Push: The Demand-Driven Approach

▶ Watch (12:02)

The standard approach pushes everything to the agent upfront: build the MCP servers, load the knowledge base, hope it works. Navakoti proposes the reverse. Assign a task. Let the agent fail. When it fails, it produces a checklist of what it needs. A domain expert fills the gaps. The agent solves the task and writes what it learned into a structured context block. Repeat across 10 or 20 work items.

This is TDD for context. Give agents problems they will definitely fail, fill those gaps, and eventually the agent becomes semi-autonomous on that domain.

Live Demo: One Cycle with Claude Code

▶ Watch (17:10)

Navakoti ran the single-incident demand cycle demo (17:10) in Claude Code using flat files as stand-ins for Confluence and Slack. The agent received one incident (notification service failing to send SMS), scored its own confidence at 1 out of 5, and listed six entities never written down anywhere. Navakoti provided a prepared answer. The agent resolved the incident and added five or six new context blocks.

The key point: you cannot know what is undocumented until an agent tries to use it. Giving agents a failing task surfaces the exact gaps.

Automating Gap Discovery at Scale

▶ Watch (26:19)

Manual cycles are too slow. The 14-incident context gap scanner run (24:12) shows the automated path. Twenty archived incidents go in. The scanner generates probes, tests them against the knowledge base, and scores each incident: documentation clean, stale, or missing. Agent confidence across the 14-incident batch rose from 1.5 to 4.4.

The output is a Kanban board sorted by criticality. If notification service gaps appear across 20 incidents, that card goes first. The agent writes its own documentation once you fill each gap.

Storing the Context Base: GitHub and the Meta Model

▶ Watch (33:22)

On storage, Navakoti is direct:

“So I have a very opinionated opinion.” — Raj Navakoti

GitHub. Multiple agents and domain experts will write to the same knowledge base, conflicts will happen, and GitHub’s PR process handles that out of the box. The repo can also sync to Confluence or Slack later. A meta model adds further value: it maps business processes to systems to APIs so an agent asked to touch a system knows which processes it will break.

Q&A

Have you used this at enterprise scale, or only toy examples? Not at full enterprise scale; he scoped down to a single team’s Jira tickets and Confluence space, finding even that boundary required narrowing further to keep one person’s domain expertise sufficient. ▶ 43:38

Have you applied it to a codebase as well as documentation? Yes, but with mixed results when combining code and Confluence because the agent could not resolve which source of truth to trust when the two conflicted. ▶ 46:50

How do you handle stale or wrong information in the context base, and what does that cost to maintain? The scanner flags documents by last-updated date and duplication, but admits it cannot detect semantically wrong information that is still current; he suggests a periodic re-scan cadence and argues per-domain token usage stays under 100k, making cost manageable. ▶ 1:00:38

Could you replace the demand cycle with meeting transcripts from onboarding calls? Possibly, if calls are recorded and transcripts are compressed before ingestion, but he was unsure whether transcripts would be cleaner than the existing documentation they already struggle with. ▶ 1:05:58

Notable Quotes

So I have a very opinionated opinion. Raj Navakoti · ▶ 33:34

it’s unreliable and it’s untested right Raj Navakoti · ▶ 8:52

how good the context is uh you’re not solving it. Raj Navakoti · ▶ 46:26

Key Takeaways

  • 40% of enterprise knowledge is tribal and never documented; pre-built retrieval layers miss it entirely.
  • Flip to a pull model: assign real work items to agents and let them demand what is missing.
  • Run archived incidents through a gap scanner to get a prioritised Kanban of context debt before going to production.