01-projects/phdata/competitive-intel/managed-agent-handoff

Open decisions - Claude partner CI agent

2026-10-22

Decisions only you can make

Each item has Ray's recommended default, labeled as a guess. Write your answers at the bottom. [[KICKOFF-PROMPT]] tells the session to read them.

1. Account: personal vs phData-owned

Guess: run it under your personal Anthropic Console account, in its own workspace. Use public data only.

Why: the agent reads only public partner information, so it needs nothing from phData. A phData-owned org would make it a phData system, which raises questions about who approves it, who pays for it and who owns its data. Do not put phData confidential information into prompts, memory stores, rubrics or the brief. That includes pipeline, deal names, quarterly figures and internal partner-program details. If phData later wants this as a team tool, move it with their approval. The ant apply files make it easy to recreate in another workspace.

2. Delivery channel

Guess: Notion page (option A in [[PLAN]] Phase 3), written into one dedicated "CI Debriefs" database. Share only that database with the Notion connection.

Why: it is the field-monitor template's built-in path, with OAuth connected in the Console, no key to handle, and a dated archive. Switch to email (option B) only if you find you do not open Notion. Until Phase 3, the brief lives in the session outputs, which you can see in the Console.

3. Cadence

Guess: weekly, Mondays at 6:00 AM Eastern Time (0 6 * * 1, America/New_York).

Why: you get it before your week starts, and it is outside the daylight-saving-time danger window. Tier promotions land on fixed dates (Jan 1, Jul 1, plus Oct 1, 2026), so trigger an extra manual run (ant beta:deployments run) on Jan 2 and Jul 2 every year (next: 2027-01-02), after refreshing the snapshot (item 9) if the directory cannot be read. Demotions happen only at the Dec 31 annual review. You asked for this cycle to be watched.

4. Watched set

Guess: the Preferred peers Slalom (you named it a big direct competitor: reported first, with a line every week even when nothing changed), IBM, NTT DATA, Thoughtworks, EPAM, Caylent, Perficient and Globant. That is 8 partners, which is a manageable weekly read. Slalom's profile (seen 2026-10-22) lists no industries, products or deployment surfaces and 0 case studies, last updated 2026-08-15, so any change there is news.

Open questions for you:

5. Web reach (networking)

Guess: limited networking with a curated host list of 25 or fewer. With limited networking, the environment's allowed_hosts also limits web_search and web_fetch. An empty list means no web results at all.

Starter list (edit it): partnerhub.claude.com, www.anthropic.com, anthropic.com, claude.com, www.claude.com, the watched set's main sites (www.ibm.com, www.nttdata.com, www.slalom.com, www.thoughtworks.com, www.epam.com, caylent.com, www.perficient.com, www.globant.com), and 3-5 press or newswire sites you trust. For section (e) (phData public image), add www.phdata.io and its blog host, phData's Snowflake and AWS partner listing pages, and the press sites above. Section (d) asks for hiring signals; job posts usually live on careers subdomains or job boards, which this list does not reach (exact-host matching). Add those hosts for Slalom and one or two others, or drop hiring signals from (d). Put the same domains in the web tools' allowed_domains. A bare host in allowed_hosts matches only that exact host, so list www. variants separately (managed-agents-environments.md).

Tradeoff: an allowlist protects you from prompt injection and noise, but the agent cannot see news that appears anywhere else. Unrestricted networking sees more, but the onboarding warns against it once the agent holds credentials (Phase 3).

6. Model and cost cap

Guess: claude-opus-5-5 at the template's low effort, with a $5.00 cap per run (budget.max_list_cost.amount: "500"), and a workspace spend limit as the backstop.

Option (you decide, not built by default): a multiagent roster with a cheaper worker (Claude Haiku 5.5, or Sonnet 5.5 when the worker needs more judgment) for the reading-heavy partner scans, with Opus doing the synthesis (managed-agents-multiagent.md, Step 2). It costs less per page read, but adds more moving parts to learn. Revisit it in Phase 5 once you know the real cost of a run.

7. Grading depth

Guess: max_iterations: 3 (the default) with the starter rubric. Raise it only if briefs regularly fail one criterion.

8. Phase 3 permission policy

Guess: auto on the Notion toolset, as the onboarding recommends for unattended runs that read outsider text. If runs pause waiting for approval, switch to a single allowed page-create tool with always_allow. Do that before Phase 4 unpauses the schedule, or add an idle webhook.

9. Directory refresh route (if web_fetch cannot read the directory)

Guess: refresh outside the agent on the promotion calendar. Re-run the Playwright scrape on a machine you control the day after Jan 1 and Jul 1 (Ray can do this), then write the new CSV into the read-write snapshot store with ant beta:memory-stores:memories update (a host-side write). The weekly agent then reports the difference. Options to test in Phase 1: curl from the sandbox to a public JSON endpoint the directory app calls (no login), or a headless browser in the sandbox (not covered by the docs; test before relying on it). Tier moves happen only on fixed dates, so this loses little; it mainly misses new Select joiners between refreshes.


My answers (Ben fills in)

  1. Account:
  2. Delivery:
  3. Cadence:
  4. Watched set:
  5. Hosts:
  6. Model / cap:
  7. Iterations:
  8. Permission policy:
  9. Directory refresh: