Episode 5 — Live
Valuestream.
Where value actually flows.
A podcast from Rick Pollick on how modern companies turn strategy into shipped software — the operating systems, delivery practices, and agentic-AI patterns moving value from the roadmap to production. New episodes Monthly.
Episode 5Episode 5 · 22 min
Context Is the Job
Why Reliable AI Agents Are a Context Problem, Not a Prompt Problem
August 31, 2026
Reliable AI agents fail on context, not prompts: what the model can actually see when it acts. A 2026 survey traced 57% of enterprise agent reliability failures to missing or inconsistent context, not the model. This episode walks the context budget, the four failure modes, and how to run context as a delivery discipline, with version control, evals, ownership, and observability.
Listen on
Episode 6
Get the next episode the morning it drops.
Same list as the blog notifications. One line per ship. No fluff in between.
The format
Three segments. One coherent show.
Every episode follows the same spine — so guests, teardowns, and solo pieces all feel like one show, and you always know what's next.
Segment
Intake
The problem or opportunity entering the stream — framing, signal, strategic intent.
Segment
Flow
How the work actually moves — operating model, platform, delivery practice, the agentic layer.
Segment
Outcome
What changed — shipped software, org behaviour, customer metrics, what we'd do differently.
Episode 5 — segment by segment
- Intake. Prompt engineering was the tutorial; context engineering is the job. A 2026 survey traced 57% of enterprise agent reliability failures to missing or inconsistent context, not the model. The fix is a delivery discipline: govern what the agent can see.
- Flow. The context budget (attention degrades before the window fills), the four failure modes (retrieval, tool-output flooding, memory that never forgets, history that never compacts), and the operating model: version control, evals, and observability, with a named owner.
- Outcome. A support agent that was confidently wrong 22% of the time drops to about 4% on the same model, by fixing what it could see: context cut from about 38k tokens to 12k, retrieval quality from 60% into the low 90s, and mean time to understand incidents measured in minutes.
Earlier episodes
Episode 4 · August 10, 2026
The Demo Isn't the Deliverable
Crossing the Production Gap From Pilot to Production
Episode 3 · July 10, 2026
The Agent Acts, You Answer
Governing the Agents You've Already Deployed
Episode 2 · June 12, 2026
Trust the Number That Hurts
The Metrics That Lie in the Agentic Era, and What to Measure Instead
Episode 1 · May 25, 2026
See It, Own It, Move It
Where Value Actually Flows in Modern Software Delivery
Want to be on the show?
Pitch a guest — or yourself.
Building an interesting value stream? Leading a transformation worth walking through? Email Rick with a one-paragraph pitch.