PlexClaw

Autonomous Agent

Settings & Knowledge

Configure PlexClaw behaviour, thresholds, and per-agent knowledge bases.

info

How Human-in-the-Loop Thresholds Work

PlexClaw uses these thresholds to decide when to act alone vs when to ask the team. Think of it as a confidence dial: turn it up and PlexClaw handles more on its own, turn it down and it asks for help more often.

Observer Mode

PlexClaw only reports. Never takes action or auto-resolves. Best for initial setup.

Supervised Mode

PlexClaw suggests actions and handles routine issues. Asks humans for new situations.

Autonomous Mode

PlexClaw handles most issues alone. Only escalates critical or high-cost decisions.

tune

Human-in-the-Loop Thresholds

Control when PlexClaw escalates to the human team vs acting alone. Lower thresholds = more human involvement.

threshold.severity_auto_resolve

Max severity Claw can handle alone. "ok" = always ask, "warning" = auto-resolve ok+warning, "critical" = never ask (dangerous)

Warning - auto-handle up to warnings, ask on critical

threshold.confidence_min

Minimum LLM confidence (0.0-1.0) before asking a human. Below this, Claw creates a request.

70%

threshold.pattern_recurrence

Number of times a pattern must recur across runs before Claw promotes it to an insight automatically.

threshold.ask_on_first_encounter

When Claw encounters a completely new type of issue with no relevant memories, should it always ask a human?

Yes

threshold.cost_action_limit_usd

Maximum cost (USD) of an action Claw can take without human approval (e.g., restarting a service, triggering a crawl).

$10

threshold.user_impact_escalation

If an issue affects more than N users, always escalate to human regardless of severity.

100

threshold.request_cooldown_hours

Minimum hours between similar human requests to avoid spamming the team.

24 hours

psychology

Agent Behaviour

Configure how PlexClaw learns, stores memories, and takes actions.

behaviour.learning_mode

Learning mode: "active" = learn from every interaction, "passive" = only learn from explicit feedback, "off" = no new memories.

Active - learn from all interactions

behaviour.autonomy_level

Autonomy level: "observer" = report only, "supervised" = can suggest actions, "autonomous" = can take low-risk actions.

Supervised - suggest actions, human approves

behaviour.memory_decay_rate

Daily decay rate for observation importance (0.0-1.0). Higher = memories fade faster.

behaviour.max_observations_per_run

Maximum observations Claw saves per run to avoid memory bloat.

behaviour.max_requests_per_run

Maximum human requests Claw creates per run.

notifications

Notifications

Control how and when PlexClaw notifies the team.

notifications.slack_enabled

Send Slack notifications

notifications.slack_critical_only

Only send Slack notifications for critical/high urgency requests.

notifications.digest_hour

Hour (0-23) to send the daily digest notification.

account_treeDecision Flow

How PlexClaw decides whether to act alone or ask for help.

visibilityIssue detected

arrow_forward

warningCheck severity

arrow_forward

psychologyCheck confidence

arrow_forward

historyCheck memories

arrow_forward

checkAuto-resolve

personAsk human