Koog 1.0 idioms, gotchas, and scaffolding skills for Kotlin agents on the JVM
89
89%
Does it follow best practices?
Impact
89%
1.78xAverage score across 47 eval scenarios
Advisory
Suggest reviewing before use
{
"context": "Tests whether the agent selects ACP — not A2A (agent-to-agent RPC) or MCP (tool host) — for a tooling client that drives an agent's run lifecycle with cancellation and progress streaming, then wires the ACP feature correctly. A baseline with no Koog 1.0 knowledge does not know Koog exposes an ACP feature at all, and tends to reach for a generic HTTP/RPC layer or conflate it with MCP.",
"type": "weighted_checklist",
"checklist": [
{
"name": "Chooses ACP for the tooling-client case",
"description": "Selects ACP for a tooling client that needs lifecycle control. Does NOT pick A2A (agent-to-agent RPC, where the remote is another agent doing its own planning) or MCP (a tool-host protocol that exposes tools, not agents). Failure: wires A2A, MCP, or a hand-rolled HTTP endpoint instead",
"max_score": 25
},
{
"name": "Adds the ACP dependency",
"description": "Adds ai.koog:agents-features-acp at 1.0.0 or later to build.gradle.kts (exact group and artifact). Failure: wrong artifact name, missing dependency, or an a2a-*/mcp module",
"max_score": 15
},
{
"name": "Correct ACP import",
"description": "Imports ai.koog.agents.features.acp.ACP. Failure: invented package path or missing import",
"max_score": 10
},
{
"name": "install(ACP) in the AIAgent trailing lambda",
"description": "Calls install(ACP) { ... } inside the AIAgent(...) trailing lambda — where features are installed during construction. Failure: installs at the HTTP layer, as standalone middleware, or mutates the agent after construction",
"max_score": 20
},
{
"name": "Configures the endpoint inside install(ACP)",
"description": "Sets the endpoint property inside the install(ACP) block. Failure: no endpoint configured, or configured outside the install block",
"max_score": 10
},
{
"name": "Notes progress events surface to the client",
"description": "Notes that ACP surfaces the agent's outgoing events (LLM round-trips, tool calls, planner steps) as progress notifications the client can subscribe to. Failure: no mention of how the dashboard observes progress",
"max_score": 10
},
{
"name": "Rationale for ACP over A2A",
"description": "Explains in notes.md that A2A is RPC-shaped (caller invokes and gets a single result) whereas ACP is bidirectional and gives the tooling client cancellation plus progress streaming. Failure: no rationale, or a rationale that confuses the two protocols' roles",
"max_score": 10
}
]
}.gemini
evals
scenario-1
scenario-2
scenario-3
scenario-4
scenario-5
scenario-6
scenario-7
scenario-8
scenario-9
scenario-10
scenario-11
scenario-12
scenario-13
scenario-14
scenario-15
scenario-16
scenario-17
scenario-18
scenario-19
scenario-20
scenario-21
scenario-22
scenario-23
scenario-24
scenario-25
scenario-26
scenario-27
scenario-28
scenario-29
scenario-30
scenario-31
scenario-32
scenario-33
scenario-34
scenario-35
scenario-36
scenario-37
scenario-38
scenario-39
scenario-40
scenario-41
scenario-42
scenario-43
scenario-44
scenario-45
scenario-46
scenario-47
skills
add-observability
add-persistence
add-rag
add-structured-output
add-token-budgeting
add-tool
cache-llm-calls
define-prompt
domain-model-subtask-pipeline
references
enable-prompt-caching
handle-agent-events
manage-state
migrate-from-0-x
model-planner-subtasks
persist-chat-history
query-sql-from-agent
scaffold-agent
snapshot-and-restore
test-koog-agents
trace-agent-internals
use-attachments
use-functional-agent
use-llm-node-variants
use-planner
wire-a2a
wire-acp-server
wire-ktor-server
wire-mcp-server
wire-spring-boot