agentdb-route

Route

Ask the Thompson Sampling bandit which approach to use for the current task.

When to use

• Task start with multiple plausible skills / algorithms
• Branching decision — A/B between approaches
• Cold start on a new task type — let the bandit explore

API

agentdb_learning_route(
  task:        <description>
  candidates?: [<skill_id> | <algo>, ...]   // omit to consider everything
  context?:    { stack, project, ... }
)

Returns: { picked, expectedReward, confidence, alternatives: [...] }

How it picks

Thompson Sampling: each candidate has a Beta(α, β) posterior over reward. The bandit samples once from each, picks the highest sample. Exploration emerges naturally — uncertain candidates get tried until their posterior tightens.

Four bandit decision points across AgentDB:

1. Pattern ranking — which historical pattern matches this query best?
2. Algorithm selection — which RL algo trains best on this task?
3. Compression tier — full / PQ8 / PQ4 / binary?
4. Skill composition — chain A→B→C or A→D→E?

The router unifies them: it returns the picked candidate AND a decisionTrace showing which decision points fired.

Use the result, then close the loop

const { picked } = await agentdb_learning_route(...)
const result = await runWith(picked)
agentdb_bandit_update(arm: picked, reward: result.reward)

The agentdb-feedback skill (this plugin) wraps the close-loop step.

Don't

• Don't second-guess the bandit on early calls — exploration is by design.
• Don't refuse the bandit's pick without recording negative reward. If you ignored a suggestion and used a different one, log that — otherwise the bandit thinks its pick "worked" because no negative signal arrived.