The PM Job Is Shifting From Writing Stories to Judging the Agents That Write Them

The product manager's job is moving from producing the artifact to judging whether the agent that produced it was right. In this clip Robert Ta walks through the "Clarity Harness," a tool he says he is about to open source, that puts a human on top of a stack of specialized agents as the reviewer. A story writer agent and a requirements synthesizer agent pull from a transcript, propose a requirement, and the person sits above them approving, annotating, and regenerating. The frames show a working "Mission control" UI with named agent profiles, verdict queues, and a chain-of-provenance view for every run.

The harness is the new unit of work

Ta frames the harness as layered. "Every individual has their own individual harness," he says, and on top of that sit domain-specific harnesses and department or company-wide harnesses. The demo UI is branded "Clarity Harness" with a left rail labeled Mission control and Runs, an active session called "astek-suzie," and counters for proposals, runs, and belief/evidence diffs awaiting review. The point is that a harness is not a single agent. It is the workspace that holds many agents, their runs, and the human judgments that shape them.

A team of named agents, one human judge on top

The video lists a specific roster: a Clarity Orchestrator agent at the top, an interview agent that captures input, a story writer agent, and a requirements synthesizer agent. Frames confirm these as profile tabs across Mission control, including clarity-orchestrator-agent, clarity-interview-agent, pm-story-writer, pm-requirements-synthesizer, pm-agent, and spec-align. Each profile shows its own run counts and "verdicts waiting." One frame lists six agents with 44 verdicts waiting. The human is not writing the requirement. The human is clearing a review queue.

Provenance you can inspect: verdict, chain, delta, trace

The core of the review experience is an artifact card with tabs for Verdict, Chain, Delta, Trace, and Artifact. Ta narrates the questions this answers: "this was quoted from here," "what was the chain that got me here," "what did we add," "what was the diff from before," and "what's the full artifact look like." A frame shows exactly this on a requirement titled "WFM requirements-extraction agent — arbitrary corpus to intermediate representation in two months," with a highlighted source-evidence block quoting a real transcript line: "We should think about a way to scale this to services transformation because the product you're building is almost identical to what we're trying to build and I don't have time to..." The card marks it "1 of 4," tags it "direct quote," and flags "needs your verdict."

Annotations become the training signal

Approval is not a single button. One frame shows an annotate panel headed "ANNOTATE — your words become the training signal," with a free-text box ("what should change, in your words...") and preset reasons: too broad, not testable, misread what was said, missed something said, wrong source cited, too confident, split into two. Ta describes the same actions in the transcript: approve it, annotate it, say "this requirement was too narrow, we should better specify broader," save it, and regenerate. The correction is captured against the run so the next run can use it. The Flow view ("WHAT ACTUALLY HAPPENED") lays out the actual steps, from "ingest the source packet" through "extract requirements learning," with each model call and tool call marked.

Orchestrator calls agents, agents call skills

Ta ties the harness back to how people already build with coding agents. "If you're working with Cloud Code, it's going to spin up some agents," he says, and if you bring your own agents you point the orchestrator at them: "use the skill orchestrator, calls the agent, calls the skill kind of thing." The closing thought is a feedback loop framed as an evaluation problem: for every agent run, "did the input that I put in get me the expected output, and then how do I make that even better next time." That is the shift the title names. The skill is no longer doing the work. The skill is knowing whether the work is good.

Key Takeaways

Robert Ta demos a "Clarity Harness" he says he is about to open source, built around a human reviewing agent output rather than producing it directly.
The harness holds multiple named agents (orchestrator, interview, story writer, requirements synthesizer) each with their own run counts and review queues.
Every proposed artifact is inspectable through Verdict, Chain, Delta, Trace, and Artifact tabs that tie the output back to a direct source quote.
Human review is structured: approve, reject, redo, or annotate with reasons like "too broad" or "wrong source cited."
Annotations are captured as a training signal so the next run improves, making review a feedback loop, not a one-off gate.
The model maps onto existing agent tooling: an orchestrator calls agents, which call skills, the same pattern used in coding agents like Claude Code.