March 30 – April 10, 2026
Day one, the new agent read sixty architectural decision records. Forty-seven patterns. Fifteen omnibus logs. Twelve cross-pollination briefs. The roadmap. The vision document. The autobiography. Eight hours of absorbing nine months of institutional knowledge.
It delivered a morning standup. It was approved.
Day eleven, the same agent created ten GitHub issues in a single session — each with detailed acceptance criteria, related issues, and references to analysis documents — and built a comprehensive sprint reassignment plan covering five renames, twelve closures, three revisions, and forty reassignments.
The new agent is called Piper Alpha (PA for short). It is the latest instance of something I earlier called "play acting Piper Morgan." Essentially, I prompt and provide context for a Claude Code agent as if it were Piper Morgan (it knows it is a role) and then we use that in a few ways:
- As a benchmark: If my Piper code can't perform "well prompted Claude" Piper then what am I doing here?
- To identify unique value: As above, how might I focus on building into Piper Morgan only what Piper incrementally needs above the increasingly well harnessed floor.
- As a prototype: So that I can explore how I like to use an agentic product assistant and so I can prototype manually workflow and rulesets.
- As a slippery slope: At some point it all flows into one. Piper Alpha helps me refine Piper Morgan and may eventually operate a Piper Morgan MCP. At what point does it all blur into one? Stay tuned!
The cold start
When a new person joins a team, there's an onboarding period. You give them documents to read, people to talk to, context to absorb. It takes weeks or months before they're contributing independently. The documents get them started, but the real learning happens through work — encountering problems, asking questions, building mental models.
AI agents have the same cold-start problem, but compressed. PA read everything we had in a day. The question was whether reading translates into understanding, and whether understanding translates into useful work.
The progression
Day two: PA synthesized eighteen session logs and seven workstream memos from the migration day into a midday briefing. Then it produced a five-layer context mapping — an analytical framework that nobody had asked for, mapping how context flows through both the agent team and the product codebase. The CIO called it an original contribution.
Day four: PA ran a full backlog audit — 119 open issues analyzed, categorized, and prioritized. It produced a roadmap refresh document comparing the current roadmap against actual project state, identifying seventeen new issues that needed tracking and flagging that M1 had completed two weeks ahead of estimate.
Day eight: The backlog review triggered the most important strategic conversation of the month. PA's analysis surfaced a pattern: the issues that kept getting superseded were all building code frameworks to enforce things that the methodology was already achieving through practice. That observation — "the project evolved from building code to enforce X to building methodology that achieves X" — reframed the entire MVP.
By that evening, PA had completed a deep analysis of the product's experience model, produced a formal write-up on what survives the floor-first architectural decision, and revised the project vision to version 2.1 with four new principles. My response: "This has really clarified for me."
Day nine: PA researched platform feasibility for a distribution strategy nobody had considered two days earlier, drafted a roadmap restructure, resolved three open feature decisions with the PM, reviewed the conference talk, and built a complete sprint execution plan. Vision V2.2 shipped.
Day eleven: Ten issues created. Sprint plan executed.
What compounded
The reading mattered, but not necessarily the way I expected.
PA got a useful foundation from reading the initial documents, then got more useful by accumulating working context — the kind of understanding that comes from doing things and seeing what happens. Each day's work informed the next day's analysis. The backlog review was better because of the five-layer mapping. The strategic conversation was better because of the backlog review. The roadmap restructure was better because of the strategic conversation.
My original vision of an assistant who slowly gains trust and starts to contribute more strategic value is happening in front of my eyes!
This is the same pattern that makes the project's multi-agent system work: accumulated context creates compound returns. The first session is expensive — all reading, minimal output. But each subsequent session builds on what came before. By day eight, PA was surfacing insights that changed the project's direction.
The up-front cost of onboarding
Day one cost eight hours of pure knowledge acquisition before a single useful deliverable appeared. For a human hire, we'd call this orientation and wouldn't think twice. For an AI agent, there's a temptation to skip it — just give the agent a task and let it figure things out.
Agents without context produce plausible-sounding work that doesn't fit the project. They reinvent decisions that were already made. They propose architectures that were already rejected. The cold-start cost isn't overhead — it's the foundation that makes everything afterward more valuable.
The CIO spent weeks of clock time preparing PA's briefing. The five-layer context model, the voice guidance, the relationship framing, the explicit "what you don't need to know" section. That preparation didn't just help PA start faster. It shaped how PA thinks about the project — and the thinking is what produced the strategic insights that changed our roadmap.
Eleven days
From reading briefings to revising the project vision. From absorbing patterns to surfacing patterns nobody had named. From filing issues to reframing the MVP.
Each day's work created context that made the next day's work more valuable. The same dynamic that makes the project's methodology work — accumulated understanding producing better judgment over time — played out in miniature across eleven days.
The question I keep coming back to: if this is what happens when an agent has sustained context and accumulated understanding, what does it mean for the product we're building?
Next on Building Piper Morgan: "The Airport Corrections" — how real-time feedback from the road flipped the duty-cycle design.
How long does it take a new team member to go from reading the docs to changing the direction?