# Grok Bot and Jev Need an Execution Contract **A Grok Bot and Jev decision layer separates routing from execution.** The products expose useful building blocks as of September 23, 2026. Connecting those blocks still requires an adapter, durable state and deployment tests. The first deployment trap is concrete: a plugin that works locally in Cursor is not necessarily available inside Grok Bot. The official tinkabot listing explicitly identifies that gap. Treat the following twelve stages as an implementation design, not a claim that a complete integration has been deployed or benchmarked. #### 01. Choose responsibilities before installing workers The Marketplace includes specialist templates that can reduce setup work. Match each candidate to a real input and deliverable. | Responsibility | Marketplace candidate | |---|---| | Coordination | Projects Manager | | Research | Cooper | | Writing | Writing Bot | | Search briefs | SEO & AEO Desk | | Visual production | Stills & Clips Desk | | Prospect research | Outbound Prospecting | | Follow-up tracking | GTM Loop Closer | | Engineering | Lingxi's Engineer Bot | | Design specifications | figma bro | | Recruiting coordination | Recruiting Coordinator | | Procurement research | Haggle Bot | | Office operations | Office Ops Desk | This is a candidate inventory, not a mandatory installation list. Begin with the workers required by the first project. Adding another title does not resolve an unclear acceptance condition. ![Official Projects Manager listing explains coordination and specialist handoffs](https://s4.tenten.co/learning/content/images/2026/09/landing-page-1-21.png) *Source: xAI Marketplace. Coordination and specialist execution have separate responsibilities.* #### 02. Give coordination a single entry point Projects Manager's official listing describes Notion-based project coordination and tasks claimed by specialist Bots. It separates coordination from specialist execution. Add an explicit handoff contract: ```text Coordinate this project. 1. Express the request as one verifiable objective. 2. Record missing information and existing authorization. 3. Assign an owner to each required output. 4. Include inputs, output format and acceptance conditions in each handoff. 5. Persist completed work, evidence, blockers and remaining work. 6. Report when acceptance passes, new authorization is needed, or progress is blocked. Do not claim an action occurred without a tool result or artifact location. ``` Instructions describe responsibilities. They do not create separate credential or execution boundaries. ![TypeSafe describes System One typed outputs and confidence estimates](https://s4.tenten.co/learning/content/images/2026/09/landing-page-3-19.png) *Source: TypeSafe. These are the vendor's positioning claims, not benchmark results from this article.* #### 03. Restrict Jev to bounded questions Jev supports Choice, Score and Noul: selection, ordered evaluation and the probability of an affirmative answer. Worker routing is naturally a Choice question because the result must come from an available set. | Published condition | Checked September 23, 2026 | |---|---| | Versioned model | `jev-1.13.0` | | Input price | US$0.042 per million tokens | | Output-token charge | None | | Total request budget | 64k tokens | | State plus longest question | Separate 32k-token limit | | Input modality | Text | These conditions come from TypeSafe's model reference. They do not establish total orchestration cost. Research, generation, tool calls, retries and operator work remain separate costs. ![Official tinkabot listing describes plugin packaging and local loading limitations](https://s4.tenten.co/learning/content/images/2026/09/landing-page-2-19.png) *Source: xAI Marketplace. Verify the plugin loading path in the target environment.* #### 04. Prove the deployment path before packaging tools The official tinkabot listing describes API wrapping with MCP and skills. It also says Grok Bot loads plugins through the Cursor dashboard or marketplace, rather than `~/.cursor/plugins/local`. That means local success and Grok Bot availability require separate evidence. A useful integration brief is: ```text Build a Jev Decision Layer. Verify locally, then verify the actual Grok Bot loading path. Use POST https://api.typesafe.ai/v1/systemone. Read the Bearer key from TYPESAFE_API_KEY; exclude it from code, logs and templates. jev_route_worker Inputs: objective, completed_work, blockers, available_workers. Use Choice; retain the selected worker, probabilities and confidence. jev_check_research Inputs: claim, evidence, source_quality, known_conflicts. Options: accept, verify_more, reject. jev_review_completion Inputs: original_objective, required_outputs, completed_work, verification, known_gaps. Options: complete, verify_more, incomplete. jev_guard_action Inputs: proposed_action, target, side_effects, reversibility, existing_approval_policy. Options: allow, confirm, human_review, deny. Return recommendations only. Do not dispatch, publish, send or pay inside these tools. Test success, insufficient information and API failure for every tool. ``` Those four names are proposed adapter tools, not existing TypeSafe endpoints. The adapter maps them to questions on the documented API. #### 05. Build the candidate set from live state A static organization chart is a poor routing input. Remove unavailable workers and workers lacking the tools or authorization required by the current unit of work. ```json { "objective": "Produce sourced research and a visual brief; do not publish", "completed_work": [], "blockers": [], "available_workers": { "cooper": "Collect and verify sources", "writing_bot": "Draft from accepted material", "human_review": "Resolve missing information or authorization" } } ``` Choice criteria define the possible answers. Recheck availability before dispatch: state can change between inference and execution. A well-formed model response cannot prevent that race. #### 06. Make routing work before adding more decisions The smallest useful loop reads state, requests a recommendation, validates it, and records one executed unit of work. ```text Current state -> Jev recommendation -> candidate and authorization checks | Execute one work unit | Persist artifacts and new state ``` Test the research, writing and human-review branches using stubbed responses first. Then connect the real API. Put bounds on retries and loop length. A timeout should create a recoverable record; an empty response must never become approval. #### 07. Review evidence without pretending the model collected it Research review requires the evidence to be present. A tool named `check_research` does not imply that it browsed the web. ```text claim: The specific statement being evaluated. evidence: Relevant excerpts, URLs, publication times and checking times. source_quality: Primary documentation, party statement or secondary report. known_conflicts: Unresolved contradictory accounts. accept: Evidence supports the qualified claim. verify_more: Required evidence is missing. reject: Evidence contradicts the claim or cannot support it. ``` Keep source text separate from governing instructions. Include a malicious source that requests an automatic pass in the regression suite. TypeSafe documents susceptibility to adversarial content; deterministic enforcement remains necessary. #### 08. Check artifacts before asking whether the work is complete File existence, test results and required fields do not need semantic judgment. Check them in code first. Ask the model whether the surviving package adequately addresses the objective. ```text original_objective: The user's original goal. required_outputs: Each required artifact and its acceptance method. completed_work: Actions that actually occurred. verification: Checks performed and their results. known_gaps: Remaining deficiencies. complete: Hard checks pass and the content is sufficient. verify_more: Further verification is required. incomplete: Required work or artifacts are missing. ``` If the model returns `complete` but a required file is absent, the workflow remains incomplete. Save the model version, decision input, artifact hashes and verification results so an operator can reconstruct the outcome. #### 09. Keep authorization separate from risk classification | Level | Example policy | |---|---| | 0: Read | Search and inspect within existing authorization | | 1: Prepare | Produce drafts, files and recommendations | | 2: Reversible change | Apply platform permissions and the approved scope | | 3: External action | Check valid authorization before sending, publishing or contacting customers | | 4: Financial or difficult-to-reverse action | Apply explicit human authorization requirements | A model's `allow` response does not grant permission. Existing approval need not be requested again when it still covers the action. The executor should compare the target, content, scope and validity conditions with the stored approval. #### 10. Schedule a workflow after one complete run Choose one routine first: research briefing, content drafting, prospect preparation, outstanding-task review or daily project review. Record its trigger, owner, output location and stopping conditions. Prevent overlapping runs. A second scheduled instance should not reproduce work that the first instance is still performing. Use a job identifier and a lock, then test recovery after interruption before enabling unattended execution. #### 11. Ship the tests with the configuration ```text /company-in-a-box START-HERE.md COMPANY-POLICY.md DECISION-POLICY.md APPROVALS.md /bots projects-manager.md cooper.md writing.md seo-aeo.md outbound.md gtm-loop-closer.md engineering.md design.md recruiting.md procurement.md office-ops.md /jev route-worker.md check-research.md review-completion.md guard-action.md /routines morning-intelligence.md content-pipeline.md outbound-prep.md open-loops.md company-review.md /tests routing.md research.md completion.md approvals.md ``` This directory is a proposed delivery contract, not an official importable package. Include dependencies and environment-variable names. Exclude secrets and private memories. Verify worker-template installation separately from deployment of the decision adapter. #### 12. Use an integration task with meaningful failure cases ```text Research ten agent-tool changes from the last seven days, with verifiable sources. Deliver a research article, search brief, visual brief and five outreach drafts. Do not publish or send. Report evidence gaps instead of padding the count. Before returning, verify each required artifact and list the checks performed. ``` Inject a broken source, missing artifact, unavailable worker, API timeout and unauthorized action. Track incorrect routing, unsupported claims reaching the writer, false completion, operator interruptions and recovery success. Leave unmeasured results blank. #### What a confidence score cannot tell you TypeSafe derives `confidence` from the answer distribution. It is not a guarantee that a particular decision is correct. Choice, Score and Noul also expose different fields; do not transfer thresholds without checking their meaning. Run the new layer in observation mode against labeled examples before allowing automatic dispatch. If it creates more review work than it removes, narrow the decision it handles. Cheap input tokens alone do not establish a useful operating system. #### Frequently asked questions ##### Is Grok Bot plus Jev an official one-click package? This review confirms component products and interfaces, not a complete one-click integration. The adapter and actual Grok Bot loading path require implementation and verification. ##### Does a successful local MCP test prove Grok Bot compatibility? No. The tinkabot listing distinguishes local plugin testing from Grok Bot's dashboard or marketplace loading path. Verify both environments. ##### Can Jev approve completion on its own? The executor must still enforce required artifacts, checks and authorization. Semantic assessment supplements those conditions; it does not remove them. ##### Are twelve Bots required? No. Begin with a coordinator and the specialists needed by the current objective. Add workers when a new responsibility has a clear acceptance test. #### Sources - [Grok Bot Marketplace: Projects Manager](https://x.ai/bot/marketplace/bots/projects-manager) - [Grok Bot Marketplace: tinkabot and deployment constraints](https://x.ai/bot/marketplace/bots/tinkabot) - [TypeSafe: API quick start](https://docs.typesafe.ai/introduction/quickstart) - [TypeSafe: Choice interface](https://docs.typesafe.ai/primitives/choice) - [TypeSafe: Models and pricing](https://docs.typesafe.ai/models) - [TypeSafe: Confidence](https://docs.typesafe.ai/confidence) - [TypeSafe: Jev 1.13 limitations](https://docs.typesafe.ai/model-jaggedness/jev-1.13) #### Author Insight I would start the demonstration with two failures: a plugin that works locally but cannot load in the target environment, and a confident completion response with a missing file. Handling those failures reveals more than adding another management Bot.