Give agents more autonomy Leading Deeper with AI — Jasmine James https://jasminejames.dev/leading-deeper/bounded-agent-work Skill source (GitHub): https://github.com/jasminelsjames/leading-deeper-ai-skills/blob/e1d2e65d3922c5e376fc0c6e8cfbb2f85722b1a7/skills/delegate-reviewable-work/SKILL.md Adaptable template based on the talk; not the original internal prompt. Version 1 / 2026-11-04 START HERE: ONE FIRST RUN Turn a small task into a reviewable change. Use the AI tool approved for your work with supplied material. Wording and layout will vary by model; look for the evidence, uncertainty and review decisions shown here. Connections and recurring runs depend on your setup. All examples use invented sources and results. They are illustrations, not reports from Jasmine’s workplace or measured outcomes. FICTIONAL INPUTS **[B1] Task and boundaries:** Draft a replacement for the registration answer in our fictional workshop FAQ. Add the approved closing time and what happens after it. Preserve the booking link and cancellation rule. Return the text here; do not edit a file or publish. Acceptance: include the timezone, keep the existing cancellation rule, and do not invent a waitlist. Event owner: Jo. **[B2] Current FAQ · registration answer:** Book at example.org/workshop. Cancel your booking at least 24 hours before the workshop. **[B3] Approved event note · 2 November 2026:** Registration closes on 12 November 2026 at 17:00 Europe/Berlin. After that time, direct registration questions to the event owner. No waitlist process has been agreed. WORKFLOW 1. Human — Give the task a clear finish: Paste the task, existing answer and approved note. Say that the result should be replacement text in the chat, with no file changes. 2. AI — Produce the change and check its scope: Compare the existing answer with the approved note, draft the replacement, and show which acceptance requirements it satisfies. Keep undecided policy visible. 3. Human — Inspect before accepting: Read the proposed text against both sources. Jo decides whether it is ready to apply and publish; the draft itself changes nothing. ILLUSTRATIVE OUTPUT Proposed FAQ replacement · draft **Replacement answer:** Book at example.org/workshop. Registration closes on 12 November 2026 at 17:00 Europe/Berlin. After that time, direct registration questions to Jo, the event owner. Cancel your booking at least 24 hours before the workshop. [B1–B3] **Acceptance check:** The proposed text includes the approved date, time and timezone [B3], retains the booking link and cancellation rule [B2], and adds no waitlist promise [B1, B3]. This is a comparison of the supplied text, not a test of a live website. **Decision and stopping point:** Jo should confirm that the event-owner contact route is useful to readers. A contact address was not supplied. No files were changed and nothing was published. WHAT TO LOOK FOR ACROSS MODELS - Replacement answer - Acceptance check - Decision and stopping point HUMAN REVIEW - Does every changed detail match the approved note, with the original booking and cancellation details intact? - Is the contact route usable before the owner applies the draft? MCP AND TOOL SETUP Local repository, file-editing and test tools are sufficient. A repository MCP is optional when the task includes remote issues or pull requests. ### Inspect and edit a bounded working area - Needed when: Required when asking the agent to implement the change; not required for a discussion-only plan. - MCP or tool examples: Local file/editor tools; An approved repository or filesystem MCP - Tools must support: Read relevant files; Edit only the agreed scope; Inspect the resulting diff - Access needed: Read and write access limited to the authorized working area. No merge, deployment or publication capability is needed. - Without the connection: Provide the relevant files and request a proposed patch for a person to apply. ### Run the acceptance checks - Needed when: Required to claim that the implementation’s checks were actually run. - MCP or tool examples: Local terminal and project test runner; An approved test-execution tool or custom MCP - Tools must support: Run the specified checks; Read their actual output; Report failures or checks not run - Access needed: Permission to run the project’s agreed checks in the approved environment. A repository MCP alone does not guarantee a test runner. - Without the connection: Provide the commands and acceptance criteria as an unrun test plan; do not report a pass. ### Read remote work context or create a review request - Needed when: Only when the task relies on remote issues or explicitly requests a pull request or merge request. - MCP or tool examples: GitHub MCP; GitLab MCP; An approved repository CLI - Tools must support: Read relevant issues, files and review context; Create an unmerged review request only if requested - Access needed: Read access to the named repository. Add review-request creation permission only when that output is explicitly requested; no merge or administration permission. - Without the connection: Paste the approved issue context and return a local diff or patch for the owner to submit. - Provider documentation: https://github.com/github/github-mcp-server ; https://docs.gitlab.com/user/model_context_protocol/mcp_server/ Use an authorized local checkout and terminal. If remote access is unavailable, provide the issue or review context as an approved excerpt and return a reviewable local change. MODEL STARTING POINT Reviewed October 30, 2026. Start with GPT-6.1 Sol (gpt-6.1-sol). A balanced starting point for a scoped code change and its acceptance checks. Start with medium reasoning effort. Move up to GPT-6 Astra (gpt-6-astra) when: Difficult design tradeoffs or repeated meaningful failures justify more reasoning and a higher cost. ALTERNATIVES - Claude Sonnet 5.5 (claude-sonnet-5-5): An alternative for the same bounded task in an approved Claude-based agent host; compare the resulting change and review effort. HUMAN RESPONSIBILITY Keep scope, acceptance criteria and the decision to merge or deploy with the responsible person. More capable does not mean more authorized. Suggested fit based on official provider documentation, not our own speed or quality benchmarks. Test on representative examples before relying on a model. A model does not connect your tools or grant access. MCP support, available tools and model selection depend on the agent host and your organization's approved setup. API PRICE REFERENCE USD per one million input and output tokens, using standard first-party API pricing. These are not subscription prices or total workflow costs. Tool calls, reasoning, retries, caching and different tokenizers affect the final bill. GPT-6.1 SOL Provider: OpenAI API model ID: gpt-6.1-sol Context: 1.05 million tokens Standard API: $2 input / $10 output per 1M tokens. Model documentation: https://developers.openai.com/api/docs/models/gpt-6.1-sol Pricing: https://developers.openai.com/api/docs/pricing GPT-6 ASTRA Provider: OpenAI API model ID: gpt-6-astra Context: 1.05 million tokens Standard API: $10 input / $50 output per 1M tokens. Model documentation: https://developers.openai.com/api/docs/models/gpt-6-astra Pricing: https://developers.openai.com/api/docs/pricing CLAUDE SONNET 5.5 Provider: Anthropic API model ID: claude-sonnet-5-5 Context: 1 million tokens Standard API: $2 input / $10 output per 1M tokens. Model documentation: https://platform.claude.com/docs/en/models/sonnet-5-5/overview Pricing: https://platform.claude.com/docs/en/about-claude/pricing Machine-readable catalog: https://jasminejames.dev/leading-deeper/model-guidance.json PROMPT Help me complete one bounded task. PREFLIGHT Verify which tools are actually available and which source locations they can access. Do not assume that naming an MCP server means it is connected. Local file and terminal tools are valid and do not necessarily use MCP. If required access is missing, ask for an approved export or excerpt; state what cannot be checked or run. INPUTS Problem / user outcome: [fill in] Authorized working area: [fill in] In scope: [fill in] Out of scope / constraints: [fill in] Acceptance checks: [fill in] Human decision to protect: [fill in] WORK Inspect the relevant context first. State your understanding of the task and any essential missing information. Implement the smallest coherent change within this scope. Run checks that actually test the acceptance criteria. If a check cannot run, report that explicitly. Keep a record of consequential assumptions and unresolved concerns. OUTPUT Return the proposed change, what it enables, the checks performed and their results, and the decisions I need to review. If I have requested a pull request, leave it unmerged. Otherwise, provide a reviewable local change. STOP Stop when the acceptance checks are complete and the change is ready for review. Do not merge, deploy, publish or contact anyone. Stop and explain if the work requires additional access, destructive changes or a material expansion of scope. Do not silently substitute a different goal. WHAT A PERSON REVIEWS - Does this solve the intended problem rather than merely satisfy the wording? - Are the checks meaningful, and do the claimed results match their output? - What did I learn by inspecting the work? What decision remains mine? EVIDENCE TO CHECK - Record whether the acceptance criteria were met and what required rework. - Track review effort separately from agent run time; faster generation alone is not a time-saving result.