Skip to Content

Features

Core Features

  • Claude Agent SDK Compatible - Familiar API patterns
  • Streaming Messages - Real-time response streaming
  • Built-in Tools - Read, Write, Edit, Bash, WebSearch, etc.
  • MCP Integration - Model Context Protocol support
  • Human-in-the-Loop - Durable execution with checkpoints
  • Skills System - Local-first skill loading with MCP fallback
  • Project Config - .isa directory for project-specific settings
  • Event Triggers - Proactive agent activation
  • Multiple Execution Modes - Reactive, Collaborative, Proactive
  • A2A Ready - Agent Card + JSON-RPC client/server adapters

Multi-Agent Features

Swarm Orchestration

  • Dynamic Handoffs - Agents decide when to hand off control via [HANDOFF: agent_name] directives
  • Streaming Events - Track agent transitions with swarm_agent_start and swarm_handoff events
  • Handoff Trace - Complete audit trail of agent handoffs
  • Max Handoffs Safety - Configurable limit prevents infinite loops
  • Shared State - State accumulates across agent handoffs
  • See Swarm Orchestration for details

DAG Task Execution

  • Dependency Ordering - Tasks specify depends_on for execution order
  • Wavefront Parallelism - Independent tasks in same wavefront run concurrently
  • Cycle Detection - Kahn’s algorithm validates DAG structure
  • Failure Cascade - Failed tasks automatically mark dependents as SKIPPED
  • Multi-Agent DAGs - Different agents execute different tasks in parallel
  • Task Status Tracking - PENDING, READY, RUNNING, COMPLETED, FAILED, SKIPPED
  • See DAG Scheduler for implementation

MultiAgent Orchestrator

  • Fixed Routing - Explicit router function controls agent transitions
  • Shared State - Mutable state passed across agents
  • Per-Agent Config - Each agent has its own skills, tools, and options
  • See Multi-Agent for details

Production Resilience

Circuit Breakers

Every service client (Model, Session, Storage, Audit, MCP) is wrapped with a circuit breaker that prevents cascading failures when a service is degraded.

from isa_agent_sdk.core.resilience import get_isa_model_circuit_breaker
cb = get_isa_model_circuit_breaker()  # failure_threshold=8, recovery=15s

Bulkhead Isolation

Separate resource pools for reasoning (model calls, default 4) and tool execution (default 8) prevent slow tools from starving model reasoning. Configurable via ISAAgentOptions.max_concurrent_reasoning and max_concurrent_tools.

Error Classification

classify_error() maps any exception to a structured ErrorContext with recovery action (retry, fallback, escalate, abort). See Error Taxonomy for the full mapping.

MCP Fallback

When MCP is unavailable, MCPFallbackService serves cached tool schemas and baseline tools (read, write, bash) so agents continue in degraded mode.

Recovery Playbooks

Automated recovery for top failure modes: model fallback (switch to cheaper model), MCP fallback (local tool cache), memory degradation (skip enrichment).

Graceful Degradation

DegradationManager lets nodes skip optional operations (HIL approval, memory enrichment, summarization) when services are degraded, rather than failing the entire agent.

Observability

OpenTelemetry Tracing

trace_service_call() wraps any service call with an OTel span. Enable via ISA_TRACING_ENABLED=true. Install with pip install isa-agent-sdk[observability].

Production Monitoring

Prometheus metrics for circuit breakers, bulkhead pools, tracing health, and MCP fallback. Includes alerting rules and SLO definitions (99.95% completion rate). Pre-flight validation via scripts/preflight_check.py.

Intent-Driven Tier Routing

SenseNode classifies intent into 5 categories and recommends a model tier:

IntentTierExamples
direct_answerFASTgreetings, simple facts
clarificationFASTshort questions
team_discoveryFAST“which team handles X?”
delegationFAST“delegate to trade”
tool_useREASONINGcomplex multi-step tasks

Enable with ISAAgentOptions(tier_routing_enabled=True). Uses existing resolve_model_for_tier() with subscription provider registry.

Lightweight Delegation

ISAAgentOptions(lightweight_context=True) skips memory, file context, and semantic tool search for A2A delegated tasks, reducing context init from ~60s to ~2s.

Advanced Features

MCP Resource Management

  • Dynamic Rule Loading: Patterns loaded from MCP resources at runtime
  • Resource Caching: Efficient pattern compilation and caching
  • Fallback Handling: Default patterns when MCP resources unavailable
  • Multi-Resource Support: PII, medical, and policy resources

Pattern Extensibility

  • MCP-Driven Patterns: Patterns defined in MCP resources
  • New PII Types: Additional violation types via resource updates
  • Severity Levels: Support for different violation severities
  • Custom Actions: Extensible enforcement action system

Medical Compliance

  • HIPAA Integration: Healthcare data protection rules
  • Medical Keyword Detection: Healthcare-specific terminology
  • Compliance Recommendations: Actionable compliance guidance
  • Risk-Based Assessment: Severity-weighted risk scoring

Integration Extensions

  • MCP Resource System: Dynamic policy management via MCP
  • External Compliance: Integration with external compliance systems
  • Audit Logging: Comprehensive compliance event logging
  • Reporting: Violation pattern analysis and reporting
  • Policy Management: Dynamic policy updates via MCP resources

Core Features

Confidence Assessment

  • Rule-based evaluation: Detects uncertainty indicators, partial response identifiers, error keywords
  • AI-enhanced assessment: Uses AI models for secondary evaluation in complex cases
  • Multi-dimensional scoring: Considers response length, language certainty, completeness
  • Dynamic thresholds: Configurable confidence threshold (default: 0.7)

Smart Error Categorization

Automatically classifies problematic responses into 6 categories:

CategoryDescriptionExample Trigger Keywords
UNCERTAINTYHigh uncertainty“not sure”, “don’t know”, “maybe”, “possibly”
INSUFFICIENT_INFOInsufficient information“not enough information”, “need more details”
AMBIGUOUS_QUERYAmbiguous query“ambiguous”, “unclear”, “multiple interpretations”
TOOL_FAILURETool execution failure“tool failed”, “execution failed”, “error occurred”
TIMEOUTRequest timeout“timeout”, “timed out”, “request expired”
TECHNICAL_LIMITATIONTechnical limitation“technical limitation”, “cannot process”, “not capable”

Graceful Failure Handling

  • Context-aware responses: Generate appropriate alternative answers based on issue type
  • User-friendly feedback: Provide constructive guidance and suggestions
  • Transparency: Honestly acknowledge limitations without providing potentially inaccurate information
  • Actionable advice: Offer specific next steps for users

References

Home