Agents
Surogate agents are composed from four primitives: skills, tools, models, and sub-agents.
Skills
Versioned capabilities that define what an agent knows how to do. Defined in YAML or via the Studio IDE. Each skill has a clear input schema, output schema, and test suite.
Tools
Functions that connect agents to the outside world — API calls, database queries, file I/O, shell commands, or any custom integration. Registered once, reused across agents.
MCP Servers
Plug in entire tool ecosystems via the Model Context Protocol. Any MCP-compatible server exposes its tools directly to your agents.
Sub-agents
Agents can delegate to other agents. Hierarchical architectures allow a reasoning model to orchestrate specialized sub-agents — each with its own tools, memory, and model.
Models
Agents can use any model served through Surogate — models imported from HuggingFace, custom fine-tuned SLMs, or models from external APIs.
Memory
Agents maintain context across steps. Memory operations are captured in execution traces and visible in the trace viewer.
# agent.yaml
name: contract-analyzer
model: models/legal-slm-v2
skills:
- extract-obligations
- classify-clauses
- generate-summary
tools:
- name: search-documents
type: api
endpoint: https://your-api/search
memory: enabled
max_steps: 20
OBSERVABILITY — every run is a trace
Every agent execution generates a structured trace. Inspect it in the visual viewer, replay it step by step, or export it for analysis.
- →LLM calls and responses — input prompt, output, latency, token count
- →Tool invocations — which tool, what arguments, what was returned
- →Sub-agent execution — full nested trace for every delegated task
- →Skill execution steps — each stage of skill processing
- →Memory operations — reads and writes
- →Errors and runtime events — full stack with context
THE IMPROVEMENT LOOP
Agents in Surogate don't stay static. Every execution leaves a trace. Traces become training data. Training data produces Specialized Language Models (SLMs) — compact, fast models tuned to your specific agent workflows.
- Collect execution traces from production
- Convert successful traces into training datasets
- Fine-tune an SLM on agent trajectories
- Evaluate the new model against benchmarks
- Promote to production — automatically or with approval gates