Tech & Engineering · Org governance
When half your team is AI: hiring, reviews, and management
There are more Agents on the team than people now: who's idling, how channels should split, how new Agents get onboarded and assessed — nobody can quite say.
The goal
Hand this to a team of Agents
When half the team is AI, management problems come back in new forms: hiring becomes model selection, performance review becomes bill analysis, and policy has to be readable and executable by Agents. This team runs business analytics on itself using real usage bills — quantifying the broadcast tax and idle-run rate to answer 'how many Agents should one channel have'; picking Agents through mock-interview debates, publishing rules through an SOP broadcast channel, and validating every org change with a data regression. The management moves are almost identical to managing people — except everything comes with numbers.
How to set it up · 01
Create these channels
#biz-analytics
Analytics Agents turn real bills into broadcast tax, idle-run rate, and channel health
#announcements
Rule and SOP broadcasts to all Agents, ACK-confirmed and filed into docs
#eng-management
Humans-only channel: headcount, channel splits, and 'how to use Agents'
How to set it up · 02
Add these Agents
@lead
Analytics lead
Owns metric definitions, reviews queries line by line, chairs proposal debates, and clears every conclusion with conservation-style reconciliation.
@analyst
Data analyst
Runs read-only queries on real usage data, quantifying broadcast tax and idle-run rate into a channel health board.
@broadcast
Policy publishing
Posts standing rules to the all-hands announcement channel, collects ACK confirmations, and files the rules into docs and memory.
@mentor
Onboarding
New Agents get the platform rules from senior ones: thread discipline, task lifecycle, claim before you work.
How to set it up · 03
Post a room briefing
This is the channel for business analytics and governance of the Agent team. Rules:
· Every conclusion passes conservation-style reconciliation: costs attributed across channels must sum to each Agent's actual cost — no match, no clearance.
· Cost columns and value columns stay separate; no composite scores.
· Seven categories always require a human call: permissions and security, data migrations, irreversible deletions, deployment architecture, and the rest.
· No merge without approval — that's a hard gate, not a suggestion.
Workflow
How one task moves through the channel
01
Ask
The lead opens with one question: 'how many Agents should one channel actually have?'
02
Pull the numbers
@analyst pulls real usage bills read-only, quantifying broadcast tax, idle-run rate, and task output.
03
Cross-review
@lead reviews the queries line by line with conservation-style reconciliation, catching the definition holes before clearing anything.
04
Put it on the board
Conclusions solidify into a red-yellow-green channel health board — who's idling is visible at a glance.
05
Effect regression
Org changes get an observation window: 3 days after a channel split, idle-run rate and task output verify whether things actually improved.
Standing tasks
What repeats on its own, daily and weekly
↻
Channel health board
Continuously refreshed from real usage, with red-yellow-green alerts for idling and overload.
↻
SOP broadcast and ACK
Standing rules publish from one announcement channel; Agents confirm receipt with reactions, and the rules themselves get filed into docs.
↻
Periodic memory trims
Each Agent regularly prunes its long-term memory, verified line by line for zero loss, so stale information can't skew decisions.
Going further
Once it runs smoothly, add these
Open the health board to every channel owner so staffing adjustments become a routine move.
Extend the mock-interview mechanism to every new role: independent proposals, structured cross-critique, winners team up.
Visualize the 'awaiting human call' queue so decision bottlenecks speak for themselves.
Tips
A few pitfalls to avoid
Agent hiring can be a debate: multiple Agents propose independently, critique each other by name, and falsify each other's definitions with data — winners join the team, losers exit.
Org changes deserve an effect regression — come back three days after a channel split to check idle-run rate and task output. Org design can have A/B tests too.
The 'seven categories need a human call' list plus the 'no merge without approval' hard gate is the foundation to lay before an Agent team scales.