| name | agenticflow-agent |
| description | Create, run, and iterate on a single AgenticFlow AI agent — one chat endpoint, one assistant, one persona. Use when the user wants a customer-facing bot, a support assistant, a single task agent, or a prompt experiment. Choose this skill over agenticflow-workforce when there's no orchestration between roles (no handoff, no coordinator → workers). Covers `af agent create/update/run/delete`, the `--patch` partial-update pattern for iteration, `af schema agent --field <name>` for nested payload shapes (including suggested_messages, mcp_clients, response_format), the `model_user_config` / `code_execution_tool_config` settings, and safe iteration loops. |
| compatibility | Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI |
| metadata | {"author":"PixelML","version":"4.0.0","license":"MIT"} |
| triggers | ["create an agent"] |
AgenticFlow Agent
A single AI agent with a system prompt, model, optional MCP tool attachments, and an optional code-execution sandbox. Use this when one chat surface + one set of rules is enough.
When NOT to use this skill
If the user needs multiple agents that hand off to each other (research → write, triage → specialist, a pre-built team template), use agenticflow-workforce instead. Don't over-engineer — a support bot with "if billing/refunds/privacy, escalate to email" is one agent, not three.
Orient first
af bootstrap --json
From the response, extract:
auth.project_id — required on agent create (server does not auto-inject for agents, unlike workforces)
auth.workspace_id
_links.workspace — surface this URL to the user right away: "Your AgenticFlow workspace is at <_links.workspace> — open it anytime to see what I'm building." Anchors a human-first mental model before any mutation
models[] — use as source of truth for model ids (don't hardcode — they change between CLI releases)
agents[] — so you don't duplicate existing work
If data_fresh: false in the response, the backend is degraded — don't mutate. Run af doctor --json --strict and fix auth/network first.
Discovery & health
af changelog --json
af context --json
af bootstrap --strict --json
af bootstrap returns an invocation block telling you the correct CLI binary to use. af bootstrap --strict exits non-zero when the backend is unhealthy, so CI/automation can abort before mutating against a degraded workspace.
Inspect payload shape before writing
af schema agent --json
af schema agent --field mcp_clients --json
af schema agent --field suggested_messages --json
af schema agent --field response_format --json
af schema agent --field update --json
The --field drilldown returns the documented shape for a single field. Use it instead of guessing.
New fields (v1.10.x)
| Field | Type | What it does |
|---|
welcome_message | string | Greeting shown on new thread |
agent_type | standard | autonomous | Default: standard |
recursion_limit | number (10-500) | Default: 100 (was 25) |
model_user_config | { temperature?, max_tokens?, max_input_tokens?, reasoning_effort? } | Fine-tune the model |
code_execution_tool_config | { enable: bool, enable_file_operations?: bool } | Python/JS sandbox |
file_system_tool_config | object | null | Enable file system tool |
attachment_config | object | null | File attachment config |
response_format | object | null | Structured output schema (JSON mode) for the final response |
knowledge | object | null | Knowledge base / RAG configuration |
skills_config | object | null | Skill pack configuration |
task_management_config | object | null | Task queue / scheduling |
sub_agents | array | Sub-agent configurations for agent teams |
plugins | array | Plugin configurations (e.g. web_search, web_retrieval) |
suggested_messages | [ { title, label, action } ] | Pre-populated prompts. NOT an array of strings — server rejects strings |
Null-rejected fields on update
These fields must be OMITTED (not sent as null) on af agent update — the server rejects null:
suggest_replies_model, suggest_replies_model_user_config, suggest_replies_prompt_template,
knowledge, task_management_config, recursion_limit,
file_system_tool_config, attachment_config, response_format, skills_config
The CLI auto-strips them even without --patch. Stripped fields are logged to stderr.
af schema agent --field update --json
Create (always preview first)
af agent create --body @agent.json --dry-run --json
af agent create --body @agent.json --json
Minimum valid payload:
{
"name": "My Support Assistant",
"tools": [],
"project_id": "<from bootstrap auth.project_id>",
"model": "agenticflow/gpt-4o-mini",
"system_prompt": "You are ..."
}
Default model change: Since CLI v1.8.1, the upstream default is agenticflow/gpt-4o-mini (was gemini-2.0-flash). GPT-4o-mini follows system prompts reliably and calls tools without refusing. Use your preferred model via --model <id>.
recursion_limit defaults to 100 (was 25). If an agent returns completed_empty, check if recursion_limit was lower.
Available models live in af bootstrap --json > models[] — always read from there rather than hardcoding a list in your logic (models ship between CLI releases). The CLI validates your model string at create time: typos fail fast with an actionable hint listing the known set. If you pass a vendor/model-name-shaped string not in the known list, it warns-but-proceeds so brand-new models work before the CLI is updated.
Run (smoke test)
af agent run --agent-id <id> --message "Test prompt" --json
af agent run --agent-id <id> --thread-id <tid> --message "continue" --json
af agent run --agent-id <id> --message "Test" --wait --timeout 60 --json
af agent stream --agent-id <id> --message "Test" --json
completed_empty (v1.8.2)
The backend sometimes returns {status: "completed", response: ""} when the agent exhausts its recursion_limit in a tool loop. The CLI now reclassifies this as:
{ "status": "completed_empty", "warning": "..." }
Exit code 2 — bash && chains halt automatically.
Remediation:
- Inspect thread messages:
af agent-threads messages --thread-id <tid> --json
- Raise recursion_limit:
af agent update --agent-id <id> --patch --body '{"recursion_limit":100}' --json
- Refine prompt to reduce loop depth
Do NOT treat completed_empty as success — the response is empty.
Get agent (aliases + fields)
af agent get --agent-id <id> --json
af agent get --id <id> --json
af agent get --id <id> --fields id,name,model --json
Iterate with --patch (the cornerstone pattern)
Never round-trip the full agent body to change one field:
af agent update --agent-id <id> --body @updated.json
af agent update --agent-id <id> --patch --body '{"system_prompt":"new prompt"}' --json
af agent update --agent-id <id> --patch --body '{"model":"agenticflow/gpt-4o-mini"}' --json
af agent update --agent-id <id> --patch --body '{"mcp_clients":[{"mcp_client_id":"<id>","run_behavior":"auto_run","tools":{}}]}' --json
The CLI auto-strips null-rejected fields (knowledge, recursion_limit, task_management_config, suggest_replies_*, file_system_tool_config, attachment_config, response_format, skills_config). Stripped fields are logged to stderr so bots don't think they cleared a field they didn't.
Attach an MCP tool provider
See the agenticflow-mcp skill for the full inspect-before-attach flow. Short version:
af mcp-clients list --name-contains "google sheets" --fields id,name --json
af mcp-clients inspect --id <mcp_id> --json
af agent update --agent-id <agent_id> --patch --body '{"mcp_clients":[{...}]}' --json
Cleanup
af agent delete --agent-id <id> --json
On errors
Every API error returns a consistent envelope with an actionable hint. Common 4xx and their hints:
- 404 → "Run the matching
list command to see available IDs" (or double-check the ID)
- 422 → "Check
details.payload for field-level errors" (pydantic returns the offending field)
- 401 → "Run
af whoami / af login"
When hint is non-empty, follow it before retrying.