Create agent
Configure an agent for export as schema v2 JSON. The classic form maps the default model and prompts;
named v2 models and graph-native memory appear in the read-only panel when a v2 template is loaded. Field names
follow the name attributes in the form; export uses the v2 contract documented in
GRAPH_DATA_MODEL.md.
Agent name becomes the generated class (e.g. SupportAgent).
Description is shown in listings. Tags help the coordinator route or filter
agents; add more with + Add tag.
For v2 exports, this row edits the default named model; additional
step models (planner, executor, etc.) are listed in the v2 models panel and preserved on export. For classic-only
agents, rows are tried in order as fallbacks. Click the row chrome to edit sampling, caching, and
transport fields under agent[model_options][].
System and user prompts may include {param} and
{prompt} (plus any parameter name you declare). The user template defaults to
{prompt}. Assistant prefill is optional chat turns added before the user message—
click Add message from the empty state, then click the row chrome (outside the inputs) for the editor. Drag by the empty chrome to reorder.
Roles cover system /
user / assistant / tool plus custom with a text field when you choose it.
Callable parameters use the same deck pattern as Model & sampling and assistant prefill: a title strip (name, type, required/optional tint), a bordered Default summary, and reorder by dragging from empty card chrome (same pointer rule as model rows—not from the × or the Default field). Rows live in the parameters bucket. The first bundled row boots with prompt required (Text); + Add parameter always inserts an empty row.
TTL controls how long responses are cached (0 = off). Cache key includes lists parameter names that participate in the cache key. Anthropic prompt caching enables provider prompt cache when supported.
Choose one return-shape preset. Selected preset is on the left, Available
(preset chips to choose from) on the right, full height. Drag a chip from the right into the selected area. The
hidden field stores the preset id (e.g. none when empty).
Retries / backoff / total timeout control transport-level resiliency. Circuit fields trip the circuit after a number of errors within a time window, then require a cooldown before retrying.
Trigger graphs, tools, and knowledge placements for the classic canvas. For v2 templates, top-level
memory/RAG fields are disabled — use mem_* nodes in graphs instead. Optional runtime
configuration under agent[harness][…]. Triggers, knowledge,
RAG, short term memory, long term memory, and tools are separate subsections
below; each can be expanded or collapsed.
Available (right) holds preset preview chips. Drag a preview into Active
to spawn a full row (unique id). Presets render as minimal or basic chips.
Rows in Active are inline (id, toggle, mode) with the graph opened from the row chrome.
Reorder Active rows by dragging empty chrome (inside editable fields you select text instead of dragging).
Drag a full row from Active onto Available to remove it from agent[harness][triggers][] — it is discarded, not listed under available.
Mirrors the Tools harness: Table, Shelf, and Library on the left;
Available on the right holds skill chips. Drag skills into a placement or back to Available.
Hidden inputs record
agent[harness][knowledge][placements][zone][] with zones
table, shelf, library, or
available.
RAG fields serialize to agent[harness][rag][…]. Chunking,
top_k, threshold, and embedding_model_id are form inputs here; ingestion and retrieval behavior are backend-defined.
Short term tier: enable. Click a chip to open settings for that style or trigger only. Drag chips between Selected and Available (drag off to remove). Configure… edits the current style or the sole selected trigger. Optional stream_capture.
Long term tier: enable, use Configure… for compaction style and triggers (same dialog as short term), and merge style (each two-bucket: Selected left, Available right, full height). Same drag behavior as short term, plus merge options for how long-term segments combine.
Zones (bench, shed, store) stack on the left; tools in Available are unstaged—drag a chip by its row to move it
between zones. After a placement, click a chip outside Available to peek a placeholder dialog (standalone preview).
Hidden names map to agent[harness][tools][zone][]. Tool-side circuit breaker fields
(errors / window / cooldown) sit below the buckets under agent[circuit_*].