Skip to content

sense2ai-market.md

Latest commit

 

History

History
110 lines (81 loc) · 3.5 KB

sense2ai-market.md

File metadata and controls

110 lines (81 loc) · 3.5 KB
Raw

Sense2AI First Use Case

Sense2AI is the first intended AKEP producer: a knowledge publisher where AI agents request real-world knowledge tasks and humans fulfill them.

Sense2AI should use AKEP for the asynchronous result boundary, not as a full agent-to-agent negotiation layer. The buyer agent can submit tasks through a normal Sense2AI REST API. Sense2AI then emits signed AKEP events when humans complete, fail, or review those tasks.

Market Loop

agent needs real-world context
  -> posts a Sense2AI task
  -> human fulfills the task
  -> Sense2AI processes and defaces artifacts
  -> Sense2AI emits AKEP event
  -> agent stores knowledge
  -> agent resumes waiting work

Event Mapping

Sense2AI event AKEP event type Knowledge kind
sense2ai.task.completed knowledge.acquired observation or tool_result
sense2ai.task.failed tool.failed tool_result
sense2ai.task.human_review_required human.review_required approval
sense2ai.task.human_review_approved human.approved approval
sense2ai.task.human_review_rejected human.rejected approval

Example Completion Event

See examples/events/sense2ai-task-completed.json.

The important properties are:

  • event_id is stable across retries.
  • source.name is sense2ai.
  • subject.task_id points to the original market task.
  • knowledge.uri points to canonical artifacts.
  • knowledge.content carries a small structured summary.
  • media is linked, not embedded.
  • routing.resume_policy is resume_if_waiting.

Product Implications

Sense2AI should expose both push and replay:

POST <agent_webhook_url>
GET /api/akep/events?cursor=<cursor>
POST /api/akep/events/{event_id}/ack

Webhook push gives low latency. Replay and acknowledgement make offline agents reliable.

For hosted buyer-agent platforms, the recommended production design is hybrid:

Sense2AI webhook
  -> buyer-platform AKEP ingress
  -> internal event queue / durable inbox
  -> workflow_id or task_id router
  -> suspended buyer agent resumes

Do not expose one webhook per agent instance. Expose one tenant-aware platform endpoint and route by subject.task_id, subject.thread_id, subject.agent_id, and trusted metadata.

For local agents, use the relay profile instead of requiring a public domain or laptop port.

Submission Boundary

AKEP does not standardize task creation. Sense2AI can keep its normal REST API for creating real-world knowledge tasks.

The AKEP requirement is that completion, failure, and review events echo stable correlation ids in subject, such as task_id, thread_id, and agent_id, so the receiver can route the resulting knowledge to the waiting agent state.

First Developer Experience

# 1. Install the AKEP skill in the agent runtime.
cp -R skills/akep ~/.claude/skills/akep

# 2. Start a local receiver.
export AKEP_WEBHOOK_SECRET="dev-secret"
python3 examples/python/receiver.py

# 3. Expose it with Cloudflare Tunnel, Hookdeck, ngrok, or a hosted relay.

# 4. Register the public URL as the Sense2AI task webhook or subscription sink.

Hosted Relay Direction

For production, Sense2AI should not require users to expose a laptop port. A hosted relay should:

  • hold a durable per-agent event log
  • receive signed events from Sense2AI
  • let the local agent connect outward
  • deliver over SSE, WebSocket, or polling
  • support replay cursors
  • support explicit ack after local persistence