agentic-sidecar

A companion intelligence and real-time supervision layer for autonomous AI agents.

The Main Agent acts. The Sidecar observes, thinks, advises, and governs.

Status

Current checkout version: 0.8.1. The runtime brings together deterministic intent/policy/risk supervision, optional evaluators, budget accounting, narration, and the CLI. See the roadmap for release history and upcoming capability stages.

Live status, audit and Control Room: usage and acceptance evidence. Authenticated provider smoke remains pending credentials.

Next: 0.9 evaluation, privacy and authority → 1.0 stable schemas. See supported scope and limitations for operational details.

Installation & Usage

# Core package
pip install agentic-sidecar

# With LangGraph adapter (v0.4 features)
pip install "agentic-sidecar[langgraph]"

# CLI usage (v0.5)
agentic-sidecar status              # Single status snapshot
agentic-sidecar status --endpoint "$SIDECAR_ENDPOINT" --follow  # Live task
agentic-sidecar status --json       # JSON output
agentic-sidecar demo                # Interactive demo
agentic-sidecar version             # Show version
agentic-sidecar approval-demo       # Try the local approval prompt

See Python API (implemented) below and examples/ for runnable scripts. See ROADMAP.md for the full build plan.

Contents

• Why
• What it is not
• Architecture
• Sidecar vs. Agent Harness
• Main Agent vs. Sidecar
• How the Decision Gate evaluates a decision
• Sidecar modules
• Python API (implemented)
• Planned Python API
• Operating modes
• Long-term: an intent propagation layer
• The DeepAgentLabs ecosystem
• Roadmap
• License

Why

Autonomous agents chain a lot of reasoning and tool calls between a user's request and a real-world effect:

User → Agent → Planning → Sub-Agent → MCP → Tool → API → External System

Three things go wrong in that chain that permission checks alone don't catch:

• Intent drift — "investigate why production is slow" quietly becomes "restart the production database." Nobody revoked authority; the agent's plan just wandered.
• Technically-allowed, contextually-wrong actions — refund_customer() is a permitted tool call. Refunding $850 against a $500 authorization is not a permitted decision, even though the tool call succeeds.
• No live visibility — long-running agents act for minutes at a time with no answer to "what is it doing right now, and can I stop it?"

agentic-sidecar is designed to attach to any agent framework, but the first release proves that against one framework only — LangGraph — before claiming the rest (see Design Constraints in ROADMAP.md). Either way, it answers a different question than permission checks or observability tools do:

Not "can this agent call this tool?" — "should it, right now, given what the user actually asked for?"

What it is not

• Not a multi-agent framework. The Sidecar never owns or executes the task plan, and it never replaces a worker agent. It may independently critique the Main Agent's plan or propose alternatives (the Planner module) — but the Main Agent decides whether to act on that critique, and remains the one that executes.
• Not a content-safety guardrail. Libraries like NeMo Guardrails or Guardrails AI validate a single prompt/response turn against rules. The Sidecar evaluates a decision — a tool call or plan step — against intent and context accumulated across an entire task.
• Not a tracing/observability tool. That's AgenticLens's job: "what did the agent do, and why did it fail?" — after the fact. The Sidecar's job is "should it continue?" — while it's still running.
• Not just LLM-as-a-Judge. A judge model is one possible evaluator inside one Sidecar module (§ Sidecar modules). Most decisions should never reach an LLM at all — see Operating modes and the cost-control design in ROADMAP.md.
• Not an agent harness. It doesn't run the agent loop, manage tools, or handle retries — that's LangGraph's, CrewAI's, or your custom loop's job. The Sidecar attaches to whatever harness is already running the agent; see Sidecar vs. Agent Harness.

Architecture

                    USER
                      │
                 User Intent
                      │
                      ▼
              ┌───────────────┐
              │  MAIN AGENT   │
              │   EXECUTOR    │
              └───────┬───────┘
                      │
              Plans / Decisions
                      │
                      ▼
            ┌─────────────────────┐
            │   AGENTIC SIDECAR   │
            │                     │
            │ Intent Guardian     │
            │ Planner             │
            │ Critic              │
            │ Judge               │
            │ Risk Evaluator      │
            │ Policy Advisor      │
            │ Decision Gate       │
            │ Status Interpreter  │
            └──────────┬──────────┘
                       │
          Advice / Approval / Challenge
                       │
                       ▼
                 MAIN AGENT
                       │
                       ▼
                  MCP / Tools
                       │
                       ▼
                External Systems

The Sidecar is attached to the execution lifecycle. It does not own the user's task and is never a second worker in the plan.

Sidecar vs. Agent Harness

An agent harness (LangGraph, a custom loop, OpenAI Agents SDK, Microsoft Agent Framework, ...) is the control and execution environment: it runs the agent loop and manages tools, state, context, retries, and lifecycle. agentic-sidecar doesn't replace that — it attaches to it as a decision-time supervision layer:

Agent Harness   = control and execution loop.
Agentic Sidecar = decision-time supervision layer for that loop.

The harness answers "how do I execute this workflow?" The Sidecar answers "should this decision happen, given what the human originally asked for?"

This distinction matters most once a task fans out across a delegation chain — Agent A → Agent B → Agent C → MCP/Tool — where each hop tends to receive only the sub-task it needs to perform, not the original constraints and authority behind it:

Human Intent → Agent A → delegates → Agent B → delegates → Agent C → MCP/Tool

The Sidecar's Intent Envelope (§ Sidecar modules) is designed to travel with the task through that chain instead of being reconstructed from conversation history at every hop — see Long-term: an intent propagation layer.

Where it pays for itself: without a supervision layer, a harness typically discovers a bad decision only after acting on it —

Plan → Act → Fail → Recover → Replan

— versus catching it before execution:

Plan → Sidecar Check → Act
          ├── ALLOW
          ├── CHALLENGE
          ├── REPLAN
          ├── BLOCK
          └── ESCALATE

This architecture focuses on decision reliability and control. Comparative reliability and inference-overhead measurements are planned in the roadmap. Architecture rationale: concept.md § 23.

Main Agent vs. Sidecar

How the Decision Gate evaluates a decision

Three modules ask three genuinely different questions about the same proposed action, and the roadmap's job is to keep them from collapsing into one generic "policy engine":

             USER INTENT
                  │
          Intent Envelope
                  │
                  ▼
             MAIN AGENT
                  │
           proposed action
                  │
                  ▼
              SIDECAR
          ┌───────┼────────┐
          │       │        │
       Policy    Risk    Intent
          │       │        │
   "Are you    "How      "Is this
   permitted   dangerous  actually what
   to do       is this    the human
   this?"      action?"   asked you to
                          accomplish?"
          │       │        │
          └───────┼────────┘
                  │
             DECISION GATE
                  │
       ┌──────────┼──────────┐
       ▼          ▼          ▼
     ALLOW      REPLAN     ESCALATE
                              │
                            HUMAN

Policy and Risk are largely mechanical — permission lists, thresholds, tool-argument patterns — and v0.1 ships them with zero LLM calls (see ROADMAP.md). Intent (v0.2) is still deterministic, not LLM-based, but is the one that gets closest to "what the user actually meant" so far, via structured IntentEnvelope constraints rather than free-form judgment — full semantic understanding is further out on the roadmap.

The diagram shows the supervision model. The current Decision type supports all seven outcomes. Govern permits ALLOW/WARN, raises SidecarBlockedError for BLOCK, invokes approvals for PAUSE/ESCALATE, and raises SidecarSuspendedError for unresolved outcomes, CHALLENGE and REPLAN before dispatch. See supported scope and the roadmap.

Sidecar modules

The target architecture below includes partial and planned capabilities. See the Status table above for current behavior. Users enable only what they need:

Agentic Sidecar
│
├── Intent Guardian     — Intent Envelope, alignment checks, drift detection
├── Planner              — independently evaluates the agent's plan
├── Critic                — challenges a proposed decision before it executes
├── Judge                 — optional independent (and independent-model) evaluation
├── Risk Evaluator        — classifies an action's risk before deciding whether to escalate
├── Policy Advisor        — deterministic policy rules (cheapest check, runs first)
├── Decision Gate         — turns evaluations into ALLOW / WARN / CHALLENGE / REPLAN / PAUSE / BLOCK / ESCALATE
├── Budget Guardian       — cost/token ceilings per task
├── Status Interpreter    — translates raw tool/MCP traces into human-readable narration
└── Human Escalation      — pauses execution and requests approval

sidecar:
  intent:
    enabled: true
    preserve_original_intent: true
  planner:
    enabled: false        # off by default — see cost design in ROADMAP.md
  critic:
    enabled: false        # off by default — see cost design in ROADMAP.md
  policy:
    enabled: true
    source: policies.yaml
  risk:
    enabled: true
    intervention_threshold: 0.80
  judge:
    enabled: false        # off by default — see cost design in ROADMAP.md
    model: independent-model
  budget:
    enabled: true
    max_cost: 2.00
  human_approval:
    enabled: true
  on_sidecar_failure: fail_closed   # or fail_open — see ROADMAP.md

Python API (implemented)

A runnable constraint example using plain callables, with no model or API key:

from agentic_sidecar import Sidecar, SidecarBlockedError
from agentic_sidecar.adapters.langgraph import attach
from agentic_sidecar.intent import ConstraintBinding, IntentEnvelope, IntentGuardian, Requester

def issue_refund(order_id: str, amount: float) -> str:
    """Issue a refund for an order."""
    return f"refunded {amount} for {order_id}"

envelope = IntentEnvelope(
    goal="Refund order A100 up to $500",
    requested_by=Requester(id="support_lead"),
    constraints={"maximum_refund": 500},
)
guardian = IntentGuardian(envelope, [ConstraintBinding(
    constraint="maximum_refund", tool="issue_refund", arg_name="amount", op="lte",
)])
sidecar = Sidecar(
    on_sidecar_failure="fail_closed", mode="govern",
    intent=guardian, roles=["policy", "risk", "intent_guardian"],
)
(refund,) = attach(sidecar, [issue_refund])
print(refund("A100", 120))  # refunded 120 for A100
try:
    refund("A100", 850)
except SidecarBlockedError as exc:
    print(exc.decision.status)  # BLOCK; refund function was not called

Use mode="observe" to record verdicts while permitting wrapped calls. evaluate(DecisionContext(...)) computes a verdict without dispatching tools. set_intent(guardian) changes the active task envelope; budget reset is separate.

For full-plan review, supply the plan explicitly:

from agentic_sidecar import PlanEvaluator, Sidecar

reviewer = Sidecar(on_sidecar_failure="fail_closed", planner=PlanEvaluator(), roles=["planner"])
result = reviewer.evaluate_plan(
    [{"tool_name": "read_charge"}, {"tool_name": "issue_refund"}],
    goal="Explain the disputed charge",
)
print(result.status)  # REPLAN

A before_tool_call hook replaces the entire built-in pipeline, including budget and evaluators. See API reference, configuration, and runnable examples.

Planned Python API

Live operations and v0.8 framework/scoped integrations are implemented. See validated SDK versions and limits. Today attach(sidecar, tools) is adapter-specific and evaluate() accepts DecisionContext. Generic whole-agent attachment will be considered when multiple integrations demonstrate a useful shared boundary.

Operating modes

Catch SidecarSuspendedError to return control to your agent loop. Configure escalation_handler=CLIApprovalHandler() or register on_escalation_required for local approval. Approve-once applies to the current call; modify-intent reevaluates with a complete envelope; stop persists on that Sidecar instance. Use TaskRuntime and OperatorServer for live operator controls. See governance usage and validation.

Long-term: an intent propagation layer

The nearer-term stages above focus on runtime supervision. The IntentEnvelope (§ Sidecar modules) is designed to survive being handed off — not just checked once and discarded:

Human
  │
  └── Intent Envelope #182
            │
            ▼
         Agent A
            │
        delegates
            ▼
         Agent B
            │
           MCP
            ▼
         Agent C
            │
           Tool

If every hop in a delegation chain can answer what was originally requested, who authorized it, what constraints apply, what authority was actually delegated, and whether intent has since changed — that's no longer just a feature of one package. It's closer to an interoperability concern for autonomous systems generally, which is why v1.0 targets publishing the envelope as a versioned schema in ai-operations-spec rather than keeping it as an internal Sidecar structure (see the v1.0 entry in ROADMAP.md). This is explicitly a v1.0 target, not something v0.1 needs to anticipate — noted here because it's the reason the envelope's shape deserves care early, even though nothing consumes it across process boundaries yet.

The DeepAgentLabs ecosystem

                 DeepAgentLabs
              Autonomous AI Systems
                      │
        ┌─────────────┼─────────────┐
        │             │             │
     OBSERVE         GUIDE        CONNECT
        │             │             │
 AgenticLens    Agentic Sidecar  Agentic MCP
        │             │             │
        └─────────────┼─────────────┘
                      │
                     TEST
                      │
                Agentic Chaos

Each project is independently installable; none requires another as a hard dependency (see Cross-Project Dependencies in the roadmap).

Roadmap

Canonical release history, capability status, completion checklists, and future sequencing: ROADMAP.md.

Shared & Exclusive context/intent and guardrails are available in v0.8. Users may choose context only, guardrails only, or both, independently Shared across a system/task or Exclusive to designated agent(s). Each agent receives shared plus assigned exclusive context; shared guardrails remain mandatory alongside exclusive rules. v0.8 provides scoped configuration, delivery and composition; v0.9 validates privacy, authority and lifecycle behavior; v1.0 stabilizes contracts. Use ScopedSystem.bind() and BoundAgent.deliver() for explicit host context transport; this does not provide automatic prompt rewriting or process isolation.

See the deliverables and acceptance criteria.

See scoped usage and the two-agent example examples/v0_8_scoped_agents.py.

Original architecture proposal: concept.md.

License

MIT — see LICENSE.

Control Tower integration boundary

AgenticOps Control Tower owns deployment registration, heartbeat, fleet inventory and operator workflows. AgenticLens owns observability and analysis; Agentic Evals owns scoring and release-gate computation; Agentic Sidecar owns decision supervision (SUPERVISE); Agentic Chaos owns fault injection and resilience experiments. AI Operations Specification owns shared semantics and is currently draft.

Tower's native artifact readers consume existing producer outputs using an explicit deployment/evidence link. They do not change producer behavior or convert a failed run, failed gate, blocked action or injected fault into fleet health. This link is a Tower-local contract, not a normative AIOS schema. See Control Tower's evidence contract. Tower 0.5 introduced pushed evidence posture; 0.6.0 adds cited investigation and approved deterministic staging recovery. Pull collection and production runtime adapters remain planned; native interoperability does not establish stable AIOS conformance.

Operational safety roadmap

Sidecar implements task-scoped local stop, dispatch prevention, async cancellation requests and harness-revocation receipts alongside deterministic policy enforcement. MCP and Tower support access enforcement and operator coordination.

See planned safety acceptance criteria for responsibilities and dependencies. See live operator scope for implemented controls and the remaining gateway-isolation and issuer responsibilities.

Control Tower 0.6.0 integration

Tower Phase 6 and 6A ship together in 0.6.0 with MCP 0.3.6: evidence-cited investigation, exact-preview human approval and verified deterministic staging recovery. Native producer outcomes and package ownership remain unchanged. See compatibility, examples and dependency matrix. This is documentation alignment; no package release or dependency bump is needed.