From Adotob Solutions
Stop Your AI Agent From Lying, Looping, or Burning Budget
Production-tested supervision patterns, heartbeat configs, eval templates, e-stop controls, and anti-fabrication protocols for autonomous AI agents.
You shipped agents. They mostly work. Sometimes they fabricate work they did not do, send the wrong email, or burn budget reasoning over an empty queue. We run an autonomous AI fleet for a 501(c)(3) nonprofit — five named agents, $3 a week in LLM spend, real customer-facing output every day — and packaged the supervision-layer patterns into a framework-agnostic kit. Plus the Receipts Governance Kit — make every agent end its work with a verifiable, rubric-graded receipt.
For OpenClaw users specifically — the hidden tax of a default setup
“$25 in 9 hours, help”
The most common OpenClaw support thread is someone watching their token budget melt while the heartbeat loops every 15 minutes on an expensive model. It’s almost always the same three problems stacked on top of each other.
On a different agent framework? The Reliability Kit below is framework-agnostic — the cost-tuning section here is OpenClaw-specific but the supervision-layer patterns are not.
Interval too short
15-minute heartbeats re-pay for the full prompt every cycle. Most provider prompt caches have a 60-minute TTL. Land inside that window and you reuse the cache instead of re-billing it.
Wrong model on the loop
Heartbeat checks don’t need reasoning. Running gpt-4o on the loop is 10× the cost of gpt-4o-mini or Haiku for the same “is there anything urgent?” check.
Always-on
Running 24 hours when you only need 12. An activeHours window of 7am–8pm cuts overnight cost to zero.
Reference setup — production box at MACONA (a nonprofit we volunteer with)
5 specialist agents running 7am–8pm ET: inbox scan, calendar awareness, content writer, editor with Playwright QA, and orchestrator. 55-minute heartbeat, gpt-4o-mini on the loop, budget cap in the config.
The free sample has the fix. The bundles are the system around it.
The free trial ships a production HEARTBEAT.md template with the 55-minute / gpt-4o-mini / activeHours config already wired in. Drop it in a workspace, point your agent at it, and your heartbeat spend stops bleeding. That’s genuinely useful on its own.
The Starter, Pro, and Enterprise bundles are what lets you safely do more than heartbeat— budget caps, e-stop controls, risk scoring, bounce circuit breakers, eval frameworks, and the constitutional rule patterns that keep an agent from freelancing with your API keys or credit card. Heartbeat tuning is step one. Running an agent that touches real systems without an e-stop is the step most people skip until the bill or the bug forces them to come back.
Pricing
Choose Your Plan
Production-tested agent configurations from real organizations. Not tutorials — battle-hardened configs.
Starter
The Reliability Kit, one-time
- ✓The Agent Reliability Kit — all 13 files
- ✓Wrapper pattern, LLM-as-judge, eval framework, RCA
- ✓Setup README + sanitized real examples
- ✓Point-in-time snapshot — no updates
Pro
All bundles + updates
- ✓All bundles: Reliability Kit + Receipts Governance Kit + more
- ✓All future bundles as we release them
- ✓Automatic updates as we iterate
- ✓New verticals (daycare, pharmacy, more)
- ✓Email support
Enterprise
Hands-on onboarding
- ✓Everything in Pro
- ✓1-hour onboarding call with founder
- ✓Custom templatization for your org
- ✓Priority email support
- ✓Early access to new bundles
Agent Reliability Kit
The supervision layer for vibe-coded agents. Wrapper pattern (kills natural-language-action fabrication), LLM-as-judge architecture, six-dimension eval framework, closed-loop bug protocol, corrections-panel retrospective, anti-fabrication brief template, and the RCA format. Sanitized real examples from a running production fleet included.
- ✓ WRAPPER-PATTERN.md — the rule that ended fabricated action reports
- ✓ LLM-AS-JUDGE-TEMPLATE.md — 4-bucket risk + 4-way decision scope
- ✓ EVAL-FRAMEWORK.md — 6-dimension declarative evals + kanban ops
- ✓ CLOSED-LOOP-BUG-PROTOCOL.md — incident → PR → RCA → monitor
- ✓ CORRECTIONS-PANEL-RETRO-TEMPLATE.md — same-day self-audit
- ✓ ANTI-FABRICATION-BRIEF-TEMPLATE.md — constraints in the brief, not at QA
- ✓ RCA-TEMPLATE.md + 5 sanitized real examples (PR diff, RCA, corrections panel, 2 eval configs)
Receipts Governance Kit
Make every agent end its work with a verifiable, rubric-graded receipt. The maintained governance moat — you get every upgraded AGENTS.mdand new config/scenario pack as we ship them, for as long as you subscribe. Framework-agnostic: drop it into any harness (Copilot CLI, Codex, opencode, ChatGPT). The intent→decompose→clarify→act→receipt loop, an honesty gate that refuses to rubber-stamp a success it can't externally verify, and cross-agent reconciliation so two agents working one objective each emit an honest receipt under a shared binding. Connects over MCP to the Receipts witness service.
- ✓ AGENTS.md — the governance block (append-safe markers), kept current
- ✓ chatgpt-project-instructions.md — same rules for ChatGPT
- ✓ GOVERNANCE-LAYERS.md — free server floor vs. adopter-owned moat
- ✓ SOP-local-harness.md + SOP-crossagent.md — single-agent + two-agent reconciliation
- ✓ receipts_emit.py — stdlib reference emitter for unattended box automations
- ✓ New scenario/config packs + AGENTS.md upgrades as we release them
Already a customer?
Access your past invoices, update payment methods, and manage your account. Enter your email and we'll send you a secure login link.
Open Customer PortalBuilt by Adotob Solutions
We run AI agents for real organizations every day — a nonprofit serving communities in West Africa and a B2B consultancy. These configs are extracted from that production experience, templatized, and packaged for you to adapt to your own organization.
Bundles include constitutional rule frameworks, cron patterns, eval schemas, and architecture docs — templatized with placeholders for your details. API keys, credentials, contact data, and proprietary scripts are not included.