Loading ToolBench
Adds AgentVeil advisory reputation and delegation signals to Paperclip workflows. Provides tools for checking agent reputation, making delegation decisions, logging interactions, evaluating teams, and generating heartbeat reports.
Server defines 5 tools with complete schemas and descriptions. Naming follows verb_noun convention (avp_check_*, avp_log_*, avp_evaluate_*, avp_heartbeat_*). All tools have descriptions (avg 150 chars, within 10-1024 baseline). All parameters are typed and described. However, output schemas are NOT documented, LLMs cannot predict response structure for downstream chaining. Error handling is absent; no recovery guidance or categorization. Missing idempotency hints on write operations (avp_log_interaction, avp_heartbeat_report). No confirmation/dry-run pattern for destructive operations. Parameter descriptions lack format constraints (e.g., DID format, score range validation rules). Composition is sound, each tool has one responsibility, but output chaining is broken by missing response schemas.
Check an agent's advisory reputation signal on AgentVeil. Returns score (0-1), confidence, and interpretation. Use before delegating tasks as a decision input.
Batch-check advisory reputation signals for all agents in a Paperclip company. Returns per-agent scores, team average, and lowest-scoring agent.
Generate an advisory status report at the end of a heartbeat cycle. Includes own reputation, velocity trend, and optional peer interaction signals.
Record an interaction signal with another agent. Log positive, negative, or neutral outcomes after task completion.
Return an advisory delegation recommendation based on AgentVeil reputation signals, with reasoning.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for avp_check_reputation, avp_should_delegate, avp_evaluate_team, avp_heartbeat_report. This breaks tool chaining, agent cannot extract score, confidence, or team_average for downstream decisions.
No error handling or recovery guidance. If AVP API returns 404 (agent not found), 401 (auth failed), or 5xx (service down), LLM receives raw HTTP error with no actionable next step. Missing error categorization (retryable vs user-fixable vs fatal).
Write operations (avp_log_interaction, avp_heartbeat_report) lack idempotency hints and confirmation pattern. Agent may log duplicate interactions on retry. No dry-run or confirmation step to prevent accidental signal pollution.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Parameter descriptions lack format constraints. 'did' parameter says 'did:key:z6Mk...' but does not specify regex, length, or validation rules. 'min_score' lacks range (0.0-1.0 is mentioned in description but not as JSON Schema min/max). LLM cannot validate input before sending.
avp_log_interaction requires 'outcome' but 'context' is optional with no default. Unclear if omitting context causes silent failure, defaults to null, or raises error. Undocumented parameter dependencies invite misuse.