| name | writing-friday-python-agents |
| description | Authoring guide for Python user agents (type:"user") on the Friday platform via the friday-agent-sdk. Covers the @agent decorator, AgentContext capabilities (ctx.llm, ctx.http, ctx.tools, ctx.stream), structured input parsing, result types, and the NATS subprocess execution model. Load when an agent.py exists in scope, when imports from friday_agent_sdk are present, when an @agent decorator is being authored or modified, or when upsert_agent was just called with type:user. Do NOT load to decide whether to author a user agent — that decision belongs in the workspace-chat agent_types rules.
|
| vendored-from | friday-platform/agent-sdk@832d539ded2c87d1a2c4ee91a2cfc8407b68eb74 |
| vendored-path | packages/python/skills/writing-friday-python-agents/ |
| vendored-version | 0.1.9 |
Writing Friday Python Agents
Friday agents are single-file Python modules that the platform spawns as native
subprocesses and communicates with via NATS. The SDK is a normal Python package
installed into your environment. There are no compile-time or WASM steps.
The mental model: your Python code runs as a normal process. All I/O that crosses
network boundaries or needs credentials routes through host capabilities the
platform provides — LLM generation, HTTP requests, MCP tools, and progress
streaming. You declare what you need in the @agent decorator, and the platform
wires it up at execution time via AgentContext.
Every Agent Looks Like This
from friday_agent_sdk import agent, ok, AgentContext, run
@agent(
id="my-agent",
version="1.0.0",
description="What this agent does — the planner reads this to decide when to invoke it",
)
def execute(prompt: str, ctx: AgentContext):
return ok({"result": "data"})
if __name__ == "__main__":
run()
Three things are non-negotiable:
- The
@agent decorator with at least id, version, description.
run() in __main__ — this is the NATS entry point. Without it, the
process spawns and immediately exits. run() checks FRIDAY_VALIDATE_ID
(registration) or FRIDAY_SESSION_ID (execution), connects to NATS, and
handles the protocol.
- Return
ok() or err() — never return raw dicts/strings.
Capabilities Quick Reference
All capabilities live on AgentContext and are always initialized. They may be
stubs in test contexts, but never None.
| Capability | Access | What it does |
|---|
| LLM | ctx.llm | Generate text or structured objects via host LLM registry |
| HTTP | ctx.http | Make outbound HTTP requests (TLS handled by host) |
| MCP Tools | ctx.tools | Call MCP server tools (GitHub, Jira, databases, etc.) |
| Input | ctx.input | Read inputFrom upstream-step docs (NOT the signal payload — that's in prompt) |
| Streaming | ctx.stream | Emit progress/intent events to the UI |
| Environment | ctx.env | Read environment variables (API keys, config) |
| Config | ctx.config | Agent-specific configuration from workspace |
| Session | ctx.session | Session metadata (id, workspace_id, user_id, datetime) |
ctx.llm — LLM Generation
response = ctx.llm.generate(
messages=[{"role": "user", "content": "Summarize this document"}],
model="anthropic:claude-sonnet-4-6",
)
print(response.text)
response = ctx.llm.generate_object(
messages=[{"role": "user", "content": "Extract key facts"}],
schema={"type": "object", "properties": {"facts": {"type": "array", "items": {"type": "string"}}}},
model="anthropic:claude-haiku-4-5",
)
print(response.object)
Pass a fully qualified model like anthropic:claude-sonnet-4-6, or set defaults
in the llm decorator param and omit the per-call model.
Optional params: max_tokens, temperature, provider_options (dict).
ctx.http — Outbound HTTP
response = ctx.http.fetch(
"https://api.example.com/data",
method="POST",
headers={"Authorization": f"Bearer {ctx.env['API_KEY']}"},
body='{"query": "test"}',
timeout_ms=30000,
)
data = response.json()
The host handles TLS, audit logging, and rate limits. 5MB response body limit.
Note: HTTP status errors (4xx, 5xx) don't raise — check response.status manually.
ctx.tools — MCP Tools
Declare MCP servers in the decorator, then call tools at runtime:
@agent(
id="my-agent", version="1.0.0", description="...",
mcp={
"github": {
"transport": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {"GITHUB_TOKEN": "{{env.GITHUB_TOKEN}}"}
}
}
},
)
def execute(prompt: str, ctx: AgentContext):
available = ctx.tools.list()
result = ctx.tools.call("create_issue", {"title": "Bug", "repo": "owner/repo"})
{{env.VARIABLE}} in MCP config references agent environment variables.
Currently only stdio transport is supported.
Do not bypass ctx.tools. Python agents must not call local MCP HTTP endpoints such as http://localhost:8002/mcp, hardcode bearer tokens, or guess provider-specific tool names. Use ctx.tools.list() to inspect the runtime tool surface and ctx.tools.call(name, args) to invoke it. Host-side tool calls are credentialed, audited, and recorded in session history; direct HTTP calls are invisible to Friday and commonly fail with unknown-tool or invalid-token errors.
If ctx.tools.call raises ToolCallError("Unknown tool ..."), list tools and fix the workspace/agent config rather than retrying a guessed name.
Platform-injected tools
The host may inject additional tools into ctx.tools beyond what you declared
in mcp={...} — workflows like human-input elicitations, memory writes,
artifact handling. These are platform features, not SDK features; their names,
arg shapes, and semantics belong to whichever platform you're running on. Call
ctx.tools.list() at runtime to see what's actually available, and consult the
platform's own docs for which tools to call. The SDK contract is just
ctx.tools.call(name, args) -> dict and ToolCallError on failure.
ctx.stream — Progress Events
ctx.stream.progress("Analyzing document...", tool_name="analysis")
ctx.stream.intent("Switching to code review phase")
ctx.stream.emit("custom-event", {"key": "value"})
Emit progress before expensive operations so the UI shows what's happening.
Structured Input Handling
A type: user agent receives input through one of two channels. Pick by how
the job is wired, not by preference:
| Channel | Where it arrives | Read with | Use for |
|---|
| Signal payload | prompt string | parse_input(prompt, ...) | Fields the trigger signal was fired with |
| Upstream step output | ctx.input | ctx.input.get("doc-id") | outputTo/inputFrom handoff from a prior step |
ctx.input does NOT carry the signal payload — a job triggered by a signal
with no upstream producer has an empty ctx.input. If you guessed a key for
ctx.input.get(...) and got nothing back, you almost certainly wanted the
signal payload — read prompt with parse_input instead.
ctx.input — Upstream step output
When the action declares inputFrom, use ctx.input rather than scraping the
prompt. Producers may compact bulky outputs into summary + artifact refs;
downstream agents dereference those refs through host capabilities instead of
asking the producer to inline large payloads.
payload = ctx.input.get("emails-result")
payload = ctx.input.artifact_json("emails-result")
emails = payload.get("emails", [])
return ok({"count": len(emails)})
Useful methods (all take an optional doc-id; omit it to operate on the full
raw input):
ctx.input.get(name, default=None) — compact input payload for an inputFrom document.
ctx.input.require(name) — same, but raises ValueError if missing.
ctx.input.artifact_refs(name) — artifact refs attached to the input.
ctx.input.artifact_json(name) — fetch via get_artifact and parse JSON contents.
prompt still exists for natural-language task instructions and backwards
compatibility, but parse_input(prompt) is not the right abstraction for
multi-step outputTo/inputFrom handoffs.
parse_input — Signal-payload extraction
from friday_agent_sdk import parse_input
data = parse_input(prompt)
@dataclass
class TaskConfig:
url: str
max_retries: int = 3
config = parse_input(prompt, TaskConfig)
Extraction strategy (3-level fallback):
- Balanced-brace JSON scan (handles nested objects)
- Code-fenced
```json blocks
- Entire prompt as JSON
Unknown keys are automatically filtered when using a dataclass schema.
parse_operation — Discriminated dispatch
For agents that handle multiple operation types:
from friday_agent_sdk import parse_operation
from dataclasses import dataclass
@dataclass
class ViewConfig:
operation: str
item_id: str
@dataclass
class SearchConfig:
operation: str
query: str
max_results: int = 50
OPERATIONS = {
"view": ViewConfig,
"search": SearchConfig,
}
config = parse_operation(prompt, OPERATIONS)
match config.operation:
case "view": return handle_view(config, ctx)
case "search": return handle_search(config, ctx)
Result Types
Always return ok() or err(). Never return raw values.
from friday_agent_sdk import ok, err
return ok({"status": "done", "count": 42})
return err("Jira API returned 403: insufficient permissions")
ok() accepts any JSON-serializable data: dicts, lists, strings, dataclass
instances (auto-converted via dataclasses.asdict).
The Execution Model — What You Cannot Do (And Why)
Your agent runs as a normal Python subprocess, so the standard library and most
installed packages work. The constraint is not a sandbox — it's that I/O should
route through host capabilities so the platform manages credentials, rate limits,
and audit logging centrally.
Recommended — use host capabilities for I/O:
ctx.http.fetch() instead of requests/httpx — host manages TLS, logging, limits
ctx.llm.generate() instead of anthropic/openai — host manages API keys, routing
ctx.tools.list() + ctx.tools.call() instead of direct API/MCP HTTP calls — MCP servers run centrally and calls remain observable
ctx.stream.progress() instead of print() — UI integration, not stdout
ctx.env instead of os.environ — only declared variables are injected
Available — Python standard library and installed packages:
json, re, base64, urllib.parse, dataclasses
collections, itertools, functools, typing
math, datetime, uuid, hashlib
pydantic if installed, numpy if installed, etc.
When in doubt: if the operation touches a network boundary or needs credentials,
use the host capability. If it's pure data manipulation, standard library or
installed packages are fine.
Getting Agents Into a Workspace
The SDK's job ends at the file: a registered @agent function plus
run() in __main__. How that file becomes a usable agent on the host
is a platform concern (Friday Studio's CLI, an in-tree dev daemon, a
deployment pipeline) — see your platform's docs for the registration
command, and do not have the agent code shell out to a daemon to
register itself.
Once the platform knows about it, reference it in workspace.yml:
agents:
- id: my-agent
type: user
The id matches the id= value from the @agent decorator. The host
resolves type: user to the registered Python agent and routes
invocations into your execute function over NATS.
The agent process is spawned per invocation by the host; it sets
FRIDAY_VALIDATE_ID (registration) or FRIDAY_SESSION_ID (execution)
plus NATS_URL, and run() handles the rest. Your code should not
open its own NATS connection.
Casing Convention
This is a subtle but real source of bugs:
- Decorator kwargs:
snake_case (Pythonic) — display_name, use_workspace_skills
- Dict values inside decorator:
camelCase (matches host Zod schemas) — linkRef, displayName inside environment dicts
- MCP config keys:
camelCase in transport config
- Result data: your choice, but
snake_case is conventional for Python agents
The bridge layer (_bridge.py) handles converting decorator metadata to
camelCase for the host automatically. You don't need to worry about the boundary
— just follow the convention above.
References
For deeper dives, read these reference files:
references/api.md — Complete API reference: every decorator param, every
context field, every method signature and return type
references/examples.md — Annotated example agents from simple (echo) to
complex (Jira multi-operation, GitHub PR operations)
references/constraints.md — Full list of constraints, casing rules,
build pipeline details, and common mistakes with fixes