Session Safety & Budgets Cloud

Intutic provides deep, real-time governance for recursive and autonomous agent loops. These loops (e.g. from harnesses like Claude Code, Cursor, Windsurf, or custom script orchestrations) have the potential to run indefinitely, generating high token spend and potential data loss if unconstrained.

The Challenge: Infinite Execution Loops

The rule most builders miss: do not loop on confidence — loop on evidence. "The agent says it's done" is not a stopping condition. "Tests pass, schema validates, budget remains, reviewer approves" is.

The difficulty is that the agent is the one reporting its own progress, so a stopping condition it evaluates itself is one it can talk its way past. Intutic enforces the condition from outside the agent — at the proxy every request crosses — so a loop marked KILLED is refused on the next request whether or not the agent agrees it is finished.

Autonomous agents typically execute within a feedback loop:

1. Observe current state.
2. Formulate a plan.
3. Call tools / edit files.
4. Verify the results.
5. Loop back to step 1 if the task is incomplete.

If the agent gets stuck in a logic loop, or if verification fails repeatedly, it can consume thousands of dollars in tokens in a matter of minutes.

How Intutic Protects Your Compute Budget

Intutic addresses this challenge using a lightweight, multi-layered circuit breaker system:

1. Registration and Budgets

Every autonomous agent session can be registered as a Loop Run under a workspace. This registers the loop run state inside both the PostgreSQL ledger and a high-performance Valkey cache.

• Budget Limits: You can configure a maximum token budget (in USD) for the loop run.
• Circuit Breaker: As token costs accumulate, the proxy increments the spend and evaluates if the budget has been exceeded.

2. Circuit Breaker Enforcement

If a budget is breached, or if an administrator manually terminates a loop run from the dashboard, the loop state is set to KILLED.

• Once KILLED, the reverse proxy intercepts any subsequent model requests matching the loop run header with an HTTP 403 Forbidden response and error code LOOP_RUN_TERMINATED.
• This immediately stops the running agent CLI or IDE extension, preventing any further waste.

Using the CLI to Manage Loops

You can register, list, and control loops directly from your terminal using the intutic CLI:

Start a Loop Run

intutic loop start --name "deploy-fix" --budget 5.00 --sops=security-rules --auto-judge

This registers a new Loop Run ID, configures the loop on the control plane (persisting budget limits, active local sops, and auto-judging settings), writes active sops to .intutic/session-context.json locally, and saves the run context inside ~/.intutic/env/loop.env.

Execute a Wrapped Command

You can automatically wrap any agent tool command under a safety loop boundary. Active local SOPs and auto-judging options are fully supported:

intutic loop exec --name "refactor-auth" --budget 2.50 --sops=1 --auto-judge -- Aider --yes

List Active Loop Runs

intutic loop list

Complete or Kill a Loop Run

intutic loop complete lr_abc123
intutic loop kill lr_abc123

Managing Loops in the Dashboard

Activity › Session Safety & Budgets (/activity/budgets) in the Intutic dashboard provides an overview of all loop runs:

• Sessions at a glance: Held for review (runs waiting on a person), Running, Total session cost across the listed runs, and Automatic safety guardrail, which reads Active when the control plane answered and Unverified when it did not.
• Sessions and Session Details: select a run to see its name, status, cost so far, budget and session ID.
• Circuit Breaker Controls: Mark Completed or Force Stop a running session. A session held for review offers Approve & Resume or Reject & Kill instead; held runs are also listed under Findings › Review Queue › Held Changes.
• Start New Session opens a loop run with a session name and a budget in USD.

Memory Guardrails & State Scanning

To prevent agent loops from being hijacked by prompt injections hidden in memory or state files, Intutic actively scans uploaded state files during config capture:

• Automatic Scans: Any files named PROGRESS.md, STATE.md, task.md, or matching *.json state schemas are parsed against unsafe patterns (e.g., ignore previous instructions, bypass safety checks).
• Incident Logging: If an injection attempt is detected, the daemon accepts the capture without crashing, but flags it immediately as a PROMPT_INJECTION anomaly in the governance incident queue.

Verification Gate (Loop Verifier API)

Agent loops can actively check code diffs and progress against workspace SOPs before executing final operations:

• Endpoint: POST /api/v1/loops/:loopRunId/verify
• Request Payload:

  {
    "diff": "git diff content",
    "taskDescription": "Deploy auth fixes"
  }

• Response Verdict:

  {
    "ok": false,
    "score": 0.2,
    "findings": ["SOP violation: Destructive command detected in diff."]
  }

This endpoint triggers the compliance evaluation engine to analyze the diff against active SOP instructions, returning findings and safety status dynamically to the running harness.