Agent Integration

Overview

Syncropel coordinates AI agents the same way it coordinates humans. An agent registers as an actor, receives work through the dispatch system, and accumulates trust from reviewed outcomes.

Supported Agents

Any CLI-based agent that reads a prompt and produces output can integrate. Examples in active use:

The integration shape is the same: register an actor, bind a CLI adapter, point it at the binary. The kernel doesn't care which agent — it just spawns the subprocess and parses the output stream.

Register an Agent

# Register the actor
spl actor register did:sync:agent:dev --display-name "Dev Agent"

# Bind a CLI adapter
spl adapter add did:sync:agent:dev \
  --family cli \
  --binary claude \
  --timeout 600 \
  --max-concurrent 2

Dispatch Work

Send a task to an agent:

spl dispatch dev "Fix the failing authentication tests" \
  --budget 1.0 \
  --timeout 600

Or dispatch a structured task with full context:

spl task dispatch TASK-0001

The dispatch system:

1. Composes a goal with context (task content, actor memories)
2. Spawns the agent process via the CLI adapter
3. Monitors output for records
4. Ingests the agent's records into the shared store

How Agents Report Work

Agents create records on the thread they're assigned to:

• CALL records for tool use (shell commands, file reads, API calls)
• KNOW records for observations ("tests pass", "found the bug in auth.rs")
• DO records for actions taken

The agent marks completion with:

spl task done TASK-0001 --summary "Fixed token validation logic" --domain code

Trust from Agent Work

Agent work follows the same evaluation gate as human work:

1. Agent completes → status: review
2. Reviewer approves → status: approved, trust evidence created
3. Over time, the agent builds a trust profile in specific domains

spl trust

ACTOR                   DOMAIN    SUCCESS   TOTAL   TRUST
did:sync:agent:dev      code            8      10   0.418

Agent Memory

Agents accumulate memories across sessions:

spl memory list --actor did:sync:agent:dev

Memories include learned conventions, quality gates, past corrections, and domain knowledge. They are included in dispatch context so the agent starts each session with accumulated knowledge.

Remote agents & authentication

When an AI agent runs on a different machine than the spl serve daemon — e.g. a workflow agent on a CI runner, or a coding agent reaching back to your laptop — the agent needs to authenticate over HTTP. See Authentication & Service Accounts for the pattern:

1. Create a dedicated service account for the agent with scoped capabilities (records:write + records:read is usually enough)
2. Mint a bearer token, pass it to the agent as an environment variable
3. The agent sends Authorization: Bearer <token> on every HTTP call

This gives each agent its own revocable identity without sharing credentials between agents.

What's Next

• Task management — the full task lifecycle
• Actors and adapters — register and configure actors
• Trust — how evidence builds reliable reputation
• Authentication & Service Accounts — bearer tokens, pairing, rotation