Walkthrough: Agent Performance Management

​

Managing AI agents like a team

Genesis gives every AI agent what every employee already has: measurable performance, a way to improve, and a career path with reviews. This walkthrough shows you the whole loop, end to end. Every screenshot is real — taken from a live Genesis environment. By the end you’ll know how measurement works, how agents improve from your corrections, and how reviews and promotions give you an evidence-backed story for your own business reviews.

​

Why this exists — “How do I know my agents are doing a good job?”

Business leaders know how to evaluate people — objectives, reviews, growth over time. Until now there was no equivalent for AI agents.

• The trust question. Before moving real work onto agents, you want evidence: how often does the agent get it right without help? Is it improving?
• The budget question. At year-end, someone has to justify what the agents did. A grading sheet an executive can read is the artifact that’s been missing.
• The improvement question. Your users correct agents every day — and those corrections vanish. You want the agent to learn from being corrected, visibly and safely.

Genesis answers all three with one employee-style model: Measure → Learn → Review & Promote.

​

Scenario 1 — Measurement: a user corrects an agent, and it counts

Here a user pushes back: the agent compared total payroll (pitchers included) against offensive output. The user says: exclude the pitchers. Genesis understands the difference between a correction (“no — that’s wrong”) and a normal follow-up (“now also add…”). Corrections become measurable signals; follow-ups never count against the agent. Exploratory conversations are never penalized for being exploratory. Nothing is asked of your users — measurement comes from conversations they’re already having.

​

The scorecard — every agent has a performance record

The scorecard turns those signals into the numbers a manager actually wants:

• First-shot success — how often the agent gets it right with no correction (here: 95.9%).
• Corrections per task — how much steering the agent needs (here: 0.05).
• Cost, feedback, volume — alongside trend charts, broken down by type of work.

Users can also thumb-up / thumb-down any answer — a direct verdict that feeds the same record. And you can back-score past months from existing history (with a cost estimate shown before anything runs), so your first report shows a trajectory, not a blank page.

​

Scenario 2 — Learning: the same correction keeps happening? Genesis proposes a lesson

When users correct an agent the same way again and again, Genesis notices the pattern and drafts a candidate lesson — with every supporting conversation attached as evidence. Nothing becomes team knowledge on its own: a person you designate reviews and approves each lesson (that’s the “Accept” button). Notice the multi-user badges — these patterns came from four different people correcting the same behavior.

​

The agent actually improves — and leads with it

Once a lesson is adopted, it reaches the agent in its next conversations. Here — one day after the pitcher-exclusion correction — the agent runs the same kind of analysis and says, unprompted: “position-player payroll only… it’s the fairer read, and it’s what I’m leading with.” The loop closes with receipts: an improvement report compares corrections-per-task before and after each adopted lesson. And when there isn’t enough data yet, it says so honestly rather than inventing a number.

​

Team knowledge — approved lessons become shared and auditable

Every adopted lesson is recorded: what was learned, which agent, which kind of work, when, and who approved it. Lessons anchored to specific data (“the results table”, “the salary table”) also land in the team wiki, where anyone can read them. When a user corrects a fact the team wiki got wrong, Genesis can even retire the outdated note and replace it automatically — with the old version preserved for audit.

​

Scenario 3 — Reviews: on a rhythm you choose, every agent gets a review

Monthly, quarterly, or a custom cycle — each agent receives a drafted performance review built entirely from recorded facts: results vs the prior cycle, strengths, recurring weaknesses and whether coaching is in place. Recurring weaknesses are tracked on their own ledger until the coaching demonstrably works — they resolve after clean cycles, and escalate if they persist.

​

Promotions — the system nominates, a person decides

Agents progress through levels — Junior → Associate → Specialist → Senior — against explicit criteria. When an agent sustains the bar, Genesis nominates it with the qualifying evidence frozen for audit. A person you designate ratifies, amends, or rejects — exactly like a human promotion. Note the honesty in the evidence table: the feedback criterion is marked skipped because there wasn’t enough feedback to judge — it never blocks a deserving agent, and it never gets invented.

​

The workforce view — levels, visible where you manage agents

Every agent’s level shows right in the Agent Center — your workforce at a glance. Here the lead agent has earned L2 · Associate while newer agents are still L1 · Junior. One promise worth remembering: levels are advisory. They inform what you choose to entrust to an agent — they never gate what an agent is permitted to do. That separation is enforced in the product, not just promised.

​

Year-end — the grading sheet for the business review

At any time — and especially at year-end — you can export a per-agent or fleet-wide grading sheet (document or PDF): level history with who decided what, cycle-by-cycle results, an executive summary. This is the answer to “justify the agents” — a defensible story of a workforce that measurably grew.

​

What we ask of you — four things, none of them heavy

1. Use your agents normally. Measurement comes from conversations you’re already having.
2. Give quick feedback when natural. A thumbs-up/down sharpens every other signal.
3. Name two owners. A lesson reviewer (minutes, periodically) and a promotion ratifier — your agents’ “manager”.
4. Pick your rhythm. Review-cycle length to match how you report internally; optionally back-score history for the first review.

Everything ships conservative by default and is tunable for your organization — cycle length, level criteria, learning-approval bars.