Teach builder to author DeepAgents skills
All checks were successful
build / build (push) Successful in 26s

This commit is contained in:
robert
2026-05-19 09:39:29 -03:00
parent e1d9eab879
commit f6c612a071
10 changed files with 857 additions and 13 deletions

View File

@@ -0,0 +1,117 @@
---
name: deepagent-agent-design
description: Design generated A2A agents around DeepAgents skills, subagents, durable workspace backends, and caller-provided LLM reasoning. Use whenever building or modifying an agent, deciding whether to add tools, creating skill bundles, wiring create_deep_agent, or avoiding shallow fake tools.
---
# DeepAgent Agent Design
Build generated agents as small A2A surfaces around a capable inner DeepAgent.
The A2A `@skill` is the user-facing API entrypoint. The inner DeepAgent does
planning, file work, skill selection, and subagent delegation.
## Default Architecture
Use this shape for non-trivial generated agents:
1. Keep one or two public A2A `@skill` methods with clear typed parameters.
2. Inside each method, read `ctx.llm`, construct `ChatOpenAI`, and build a
DeepAgent with `create_deep_agent`.
3. Pass `backend=ctx.workspace_backend()` so DeepAgents file tools write to the
caller's durable workspace instead of LangGraph state.
4. Pass `skills=[RUNTIME_SKILLS_ROOT]` when the generated project includes
`skills/<skill-name>/SKILL.md`.
5. For custom subagents, include a `skills` field on each subagent definition.
The general-purpose subagent inherits main skills, but custom subagents do
not.
DeepAgents skills are progressive-disclosure folders:
```text
skills/
skill-name/
SKILL.md
references/...
scripts/...
assets/...
```
`SKILL.md` must start with YAML frontmatter containing `name` and
`description`. The description is the trigger; include exactly when the skill
should be used. Keep detailed reference material in separate files that are
linked from `SKILL.md`.
## Runtime Skill Seeding
Project `skills/` files are source files in the agent image. DeepAgents loads
skills from its backend, so seed those files into the invocation workspace
before `create_deep_agent`:
```python
from pathlib import Path
RUNTIME_SKILLS_ROOT = "/outputs/{{ agent_name }}/.deepagents/skills/"
def _seed_runtime_skills(backend: Any) -> None:
root = Path(__file__).parent / "skills"
if not root.exists():
return
uploads: list[tuple[str, bytes]] = []
for path in root.rglob("*"):
if path.is_file():
rel = path.relative_to(root).as_posix()
uploads.append((RUNTIME_SKILLS_ROOT + rel, path.read_bytes()))
if uploads:
backend.upload_files(uploads)
```
Then:
```python
backend = ctx.workspace_backend()
_seed_runtime_skills(backend)
return create_deep_agent(
model=model,
backend=backend,
skills=[RUNTIME_SKILLS_ROOT],
tools=[small_deterministic_tool],
subagents=[...],
system_prompt=SYSTEM_PROMPT,
)
```
Because this uses `/outputs/...`, the write fits the normal A2A workspace
grant. If the agent uses `ctx.workspace_backend()`, also declare
`workspace_access = WorkspaceAccess.dynamic(...)` on the A2A agent class.
## Tool Design Rules
Do not make a long list of fake tools that just return canned text or call the
LLM again. Prefer:
- DeepAgents skills for procedural/domain knowledge the LLM should apply.
- Subagents for independent research, analysis, review, or transformation
lanes that benefit from another LLM pass.
- Small deterministic tools only for exact operations: parsing, validation,
API calls, math, file conversion, sandbox commands, database queries.
If a tool's body is mostly prompt text, it should usually be a DeepAgents
skill instead. If a tool needs judgement, route that judgement through the
inner DeepAgent or a subagent, not a hard-coded placeholder.
## Skill Authoring Workflow
When the generated agent needs reusable workflow knowledge, first call
`write_agent_skill` to create `skills/<skill-name>/SKILL.md` and any support
files. Then wire `agent.py` to seed and pass those skills.
Use concise skill descriptions with concrete trigger phrases. Example:
```yaml
---
name: market-research
description: Plan and run multi-source market research, delegate subtopics to subagents, save findings files, and synthesize cited reports. Use for competitive analysis, market maps, customer research, or current market questions.
---
```
Keep A2A public schemas stable and simple. Put rich autonomous behavior behind
the inner DeepAgent, not in a pile of public endpoints.