100 Agents in 100 Days

Series in progress

This page is the planning catalog for the series. Daily tutorials are being written and will be linked here as they publish. Browse the use cases below to see what’s coming.

Each day adds one focused tutorial or example: runnable demos, integration patterns, and production-minded snippets. The catalog below is the backlog — daily tutorials will be chosen from this list and linked here as they ship.

Why Astrocyte — the honest filter

Astrocyte’s value proposition is AI long-term memory. Long-term memory earns its infrastructure in three shapes:

1. One user, many agents. You already work across Claude, Codex, Gemini CLI, OpenCode, GLM — and next quarter’s tool. Memory locked inside one vendor’s product dies when you switch surfaces. Astrocyte is the memory layer all your agents share: retain from one, recall from any other.
2. Many users, shared memory. Teams and organizations where each person’s agents read from and contribute to institutional memory — with banks, access control, PII policy, and audit. This is Astrocyte in server mode.
3. Corpus-scale memory. When “memory” is 84 SEC filings or three years of support tickets — larger than any context window or file tree — retrieval quality is the product. This is what the LoCoMo / LongMemEval / FinanceBench benches measure.

And the anti-pattern, stated plainly: one user talking to one AI does not need Astrocyte. A local CLAUDE.md or a folder of markdown notes is simpler and works. Tutorials in this series must never pretend otherwise — every tutorial has to justify itself against the markdown baseline.

The two tests every tutorial must pass

• Accumulation test. Does memory compound over sessions? Stateless Q&A (“what does this error code mean?”) fails — nothing worth remembering tomorrow.
• Markdown test. Why aren’t local files enough? Passing answers: the memory is consumed by more than one agent (shape 1), more than one user (shape 2), or is too large for files-in-context (shape 3).

The signature proof in every tutorial: retain from one surface, recall from another — a second agent CLI, a second user, or a second machine. If a tutorial can’t stage that moment, it doesn’t belong in the series.

Stack context: The v0.11.x line builds on the v0.9.x feature surface (wiki compile, time travel, gap analysis, entity resolution, AGE graph adapter) and the v0.10.x retrieval-quality work (HyDE, observation consolidation, agentic reflect, fact-level causal links, link expansion). v0.11.x adds schema-per-tenant isolation, mental models as a first-class SPI, intent-driven reflect routing, and the four-preset benchmark ablation matrix. Tier 1 storage adapters (astrocyte-postgres, astrocyte-age, etc.), optional astrocyte-gateway-py, optional recall_authority, and ingest connectors (Kafka, Redis, GitHub poll) all ship out of the box. See CHANGELOG.md for release notes.

What to expect

• Agent patterns (tool use, multi-step flows) with durable memory via Astrocyte — always demonstrated across at least two agents or two users, never one-agent-in-a-vacuum
• Pairings with the reference REST gateway, poll ingest, or an API edge where the tutorial needs a server-side hook
• Use-case profiles (support, coding, personal, research) as the starting policy configuration for each tutorial
• Occasional deep links into the plugin SPIs, agent framework middleware, and design docs

Suggested opening week

The first five days cover all three memory shapes, and each day reuses the previous day’s setup:

Day	Tutorial	Shape it proves
1	#6 Personal CRM — retain via a Pydantic-AI script, recall from Claude Code over MCP	Cross-agent, smallest possible signature proof
2	#24 Side-project coding helper — Claude Code and Codex sharing one repo’s decisions	Cross-agent, for the developer audience
3	#5 Personal second brain — notes, PDFs, and saved links via the Document Engine	Corpus + cross-agent
4	#30 Club / PTA secretary — two members’ agents, one institutional memory	Shared, no enterprise infra needed
5	#46 Internal KB Q&A with citations — server mode, research profile	Shared + corpus, first enterprise day

Use case catalog (before Day 1)

This section is the backlog for the series—not a commitment to order. Items are grouped so we can rotate across domains (productivity, home, commerce, support, risk, engineering, etc.) and cover different agent shapes (single-turn Q&A, tool loops, scheduled jobs, human-in-the-loop). Each group is tagged with the memory shape that justifies it against the markdown baseline: cross-agent (one user, many AIs), shared (household or team), or corpus (too large for files-in-context).

Catalog provenance: rewritten 2026-07 against the memory-shape filter. Rows that were stateless or one-shot were merged into compounding neighbors or cut; every surviving row states the memory that accumulates.

Consumer use cases (individuals & households)

Consumer tutorials are built cross-agent by default: the same memory bank consumed from at least two different tools — e.g. retained via Claude Code over MCP, recalled from a Gemini CLI session, or vice versa. Household groups add the shared shape: two family members’ agents reading one bank.

Productivity & communication — cross-agent

1. Daily briefing whose source preferences and read history live in one bank consumed by your phone assistant and your desktop agent alike
2. Inbox triage that learns your priority signals over weeks — sender weights, thread patterns, unsubscribe candidates
3. Calendar negotiation grounded in accumulated scheduling preferences (protected hours, timezone habits, meeting-length tolerances)
4. Meeting notes → decisions, owners, and follow-up drafts — a decision log both you and your teammates’ agents can recall (shared)
5. Personal “second brain”: recall across notes, PDFs, saved links, voice memos, and long-thread summaries — one corpus, every agent (corpus, flagship)
6. Personal CRM: birthdays, last touchpoints, gentle nudge suggestions — retained by one agent, recalled by any other (flagship — Day 1)
7. Writing assistant whose voice examples and style canon follow you from Claude to Gemini to whatever ships next

Money, shopping, subscriptions — cross-agent · shared (household)

8. Spending categorization and “where did my money go?” Q&A over months of accumulated exports
9. Subscription audit: renewal dates, duplicate tools, downgrade prompts — plus usage patterns compiled for bill-negotiation prep
10. Comparison shopping grounded in your accumulated buying criteria and past purchase outcomes (price sensitivity, warranty weight, return-policy lessons)
11. Household document vault: receipts, warranties, leases, and consumer contracts — “what’s covered, until when, and what did we agree to?” for both partners’ agents (corpus)

Home, mobility, logistics — shared (household)

12. Relocation as shared project memory: utilities, mail, registrations — two people’s agents working one checklist over months
13. Home maintenance from appliance manuals (corpus) plus a seasonal maintenance log that compounds year over year
14. Home inventory for insurance with photo-to-line-item assist — grows with every purchase
15. Car maintenance log: service history, part numbers, and mechanic notes accumulated across years and drivers
16. Smart-home routine suggestions grounded in household routine memory both partners’ assistants share

Food, health-adjacent, wellness — shared (household) (stay within non-clinical boundaries: planning and education, not diagnosis)

17. Meal planning from pantry, accumulated diet preferences and macro goals, and time budget — one bank, two cooks
18. Recipe scaling, substitutions, and grocery list consolidation — a house recipe memory that remembers what worked
19. Fitness programming from equipment, schedule constraints, and a training log (volume, soreness notes, trends) that compounds
20. Sleep, stress, or reflection journaling with pattern surfacing — logged from your phone, analyzed from your desktop (no medical claims)

Learning, career, creativity — cross-agent

21. Tutor with spaced prompts grounded in your materials and your progress history — spaced repetition is long-term memory (flagship)
22. Language practice partner whose corrections are scoped to your accumulated learner level
23. Resume and job-description alignment from a living experience-bullet bank — cover letters and interview prep draw from the same memory
24. Side-project coding helper: Claude Code and Codex working the same repo, sharing decisions, conventions, and dead ends through one bank (flagship — the developer demo)
25. Creative writing partner with a world bible / canon that follows you across every tool you draft in
26. Reading list curator: accumulated saved articles ranked into “read next” (corpus)

Family, social, local life — shared (household)

27. Kids’ activity and weekend planning within accumulated constraints (ages, budget, travel radius, what flopped last time)
28. Gift coordination: preferences and occasions accumulated over years — two parents’ agents that never buy the same gift twice
29. Trip itinerary drafts grounded in past-trip learnings (pace tolerance, accessibility needs, “never book the 6am flight again”)
30. Club or PTA secretary: agendas, minutes, and decisions — institutional memory of a small org, no enterprise setup required (flagship — shared)
31. Wedding or large event task graph as shared project memory: couple, planner, and family agents on one dependency board

Hobbies & specialized leisure — cross-agent

32. Gardening planner whose season logs compound year over year by hardiness zone
33. Pet care schedules and vet visit prep from accumulated past notes — every family member’s agent knows the dog (shared)
34. Maker / electronics project memory: BOMs, supplier lessons, and design decisions shared between your CAD assistant and your chat agent
35. Photography curation over a photo-metadata corpus: albums, captions, duplicate detection hints (corpus)

Accessibility & life admin — cross-agent · shared (caregiver)

36. Form-filling companion drawing on a PII-governed profile bank (user verifies every field; showcases the support profile’s redaction policy)
37. Immigration or benefits checklists as structured trackers spanning months of process and multiple helpers’ agents (informational)

Enterprise use cases (organizations)

Enterprise tutorials run Astrocyte in server mode — shape 2 by construction: multi-user banks, access control, PII policy, and audit from the use-case profiles. Many also exercise corpus scale (ticket history, contract repositories, runbooks) where files-in-context breaks down. The signature proof here is two users, one memory: one teammate’s agent retains, another teammate’s agent recalls.

Customer-facing: support, success, sales

38. Tier-1 support triage with macros, entitlements, and escalation rules
39. Technical support co-pilot grounded in runbooks and known-error DBs (corpus)
40. Customer health digests for CSMs from tickets, usage, and CRM notes
41. Sales qualification assistant: ICP fit, stakeholder map, next-best action
42. Outbound personalization within brand and compliance templates
43. RFP / security questionnaire first drafts from a canonical answer bank that improves with every submission

Internal IT, workplace, HR

44. IT service desk resolutions suggester with change-risk hints
45. Employee onboarding bot across systems-of-record handoffs
46. Internal knowledge base Q&A (Confluence, Notion, SharePoint) with citations (corpus, flagship)
47. HR policy Q&A with locale and employee-type variants
48. Benefits navigator with life-event checklists
49. Leave, overtime, or timesheet anomaly explainer grounded in the policy corpus and precedent decisions

Engineering, quality, platform

50. CI failure summarizer: likely owner, flakiness history, similar incidents
51. Release notes from commits/PRs mapped to customer impact
52. On-call handoff summarizer across alerts, deploys, and bridges — institutional memory of incidents
53. Code review comment generator with repo style rules every developer’s agent shares (cross-agent too)
54. Test ideas from user stories, grounded in the team’s accumulated risk-framing and past escaped-defect patterns
55. API doc drift checker with spec history: what changed, when, and what broke last time
56. Developer onboarding navigator for monorepo topology and service owners

Security, risk, compliance

57. Security alert enrichment: MITRE mapping, asset context, playbook step-one
58. Phishing or suspicious-message triage assistant for SOC L1 — campaign patterns accumulate
59. Vendor due diligence checklist from uploaded SOC2 / SIG questionnaires (corpus)
60. Data-classification and sensitivity tag suggestions on document batches
61. Audit narrative prep: control mapping to evidence inventory lists
62. Access review campaign assistant (overprovisioned role surfacing—not decisions)

Finance & procurement (accounting / finance sign-off for binding numbers; counsel for commitments)

63. Month-end close checklist and flux-comment first drafts — flux narratives compound quarter over quarter
64. AP exception handling: three-way match breaks explained with next step, grounded in past exception patterns
65. Procurement bid tabulation and non-binding award recommendation memo from house templates and award history

Legal department & legal operations (privilege, work product, retention, and local bar rules apply—GC-approved AI policy; no autonomous legal advice)

66. Contract review prep: clause checklist hits, defined-term consistency, and diffs vs house templates (escalation to counsel)
67. Legal hold tracking Q&A: custodian lists, scope narratives, and collection status read-only from approved systems
68. NDA / MSA / order-form playbook assist: flagged deviations, missing exhibits, and fallback language options for lawyer edit
69. Data-processing addenda (DPA / SCCs) cross-check against data map, subprocessors, and transfer mechanisms
70. Matter intake: conflicts pre-screen questionnaire assembly from client / counterparty facts—requires conflicts partner clear
71. eDiscovery logistics: custodian interview prep; preservation notice shells with merge fields—no send without review
72. Privilege log drafting support: suggested descriptions from metadata + reviewer tags—no privileged conclusions
73. Deposition / hearing / filing deadline calendars from rules sets and docket extracts—human verifies in jurisdiction
74. Regulatory or agency comment-letter outline + citation pull from docket IDs and internal position memos
75. Trademark clearance knockout memo scaffolding over an accumulating search-history corpus—not a clearance opinion
76. Patent prior-art landscape memo draft over the team’s accumulated classification and query corpus—paired with searcher / prosecutor workflow
77. Outside-counsel invoice review: phase-code compliance, block-billing patterns, staffing mix vs guidelines
78. Board / committee resolution packs: cross-check duties charter, prior minutes, and required approvals
79. Commercial claims intake: limitation-period calendar hints from rough dates and governing law—verified by counsel
80. Entity management: officer/director / registered-agent delta reports vs filings (jurisdiction-specific tooling)
81. Policy / attestation campaigns: plain-language FAQ over codes of conduct, insider trading, gifts—non-binding employee Q&A
82. Sanctions / export-classification checklist triage for product specs—routed to trade-compliance + counsel
83. Contract lifecycle: renewal / termination notice windows, price ramps, and auto-renew flags from CLM extracts
84. Litigation / investigation budget scenarios from historical phase spend and staffing models—planning aid only

Operations, supply chain, field

85. Logistics exception agent: ETA slips, alternate routes, customer comms drafts — exception patterns accumulate
86. Inventory reorder hints under MOQ, lead time, and storage constraints
87. Quality inspection summaries from structured checklists and free text
88. Field service notes → structured work orders and parts lists — field and office agents share one memory
89. Manufacturing SOP Q&A for floor tablets with document version pins (corpus, flagship)
90. Facilities ticketing triage: category, priority, likely vendor skill

Data, analytics, product

91. Data catalog natural language queries with lineage snippets (corpus)
92. “Why did this metric move?” analyst grounded in past investigations — guardrails on causality language
93. Experiment readouts: segments, novelty, and recommended follow-ups
94. Product requirement ambiguity checker against the team’s accumulated acceptance-criteria standards
95. Churn risk narrative for accounts with feature-usage surrogates

Industry-flavored (each needs explicit policy + human gates)

96. Prior authorization checklist assistant for admin staff (not clinical decisioning)
97. Pharmacovigilance case intake structuring from free-text reports (validated workflow)
98. Insurance FNOL triage: missing fields, fraud sketches, human adjudication
99. Loan origination document completeness tracker
100. Wealth advisor meeting prep from CRM notes (restricted retention and disclaimers)

Marketing, comms, design systems

101. Campaign brief distiller across channel plans and brand guidelines (corpus)
102. Crisis monitoring digest from feeds with source diversity requirements
103. Content localization reviewer — the glossary and banned-claims list is itself shared memory every reviewer’s agent consults
104. Web accessibility findings explainer with per-component findings history (paired with tooling output)

Strategy & leadership

105. Executive briefing from departmental weeklys with conflict detection
106. OKR drafting assistant aligned to historical commitments and dependencies — institutional memory as input
107. Architecture Decision Record (ADR) co-author from meeting notes — ADR corpus + whole team + every developer’s agent (flagship — all three shapes)
108. Post-merger integration task master across workstreams (dependency graph)

---

How this turns into 100 days

• Pick one row per day (or merge two thin rows into a richer tutorial). The catalog is intentionally >100 numbered rows (consumer 1–37; enterprise 38–108) so we can swap domains when a dependency lands (e.g., gateway operations, poll ingest, identity plugins).
• Every tutorial must pass the two tests (accumulation + markdown) and stage the signature proof: retain on surface A, recall on surface B. A row that can’t stage it gets merged into one that can, or cut.
• Tag each tutorial twice — by agent shape (reactive chat, scheduled job, tool loop with approvals, multi-agent handoff, offline batch) and by memory shape (cross-agent, shared, corpus).
• Enterprise days should link to identity and policy docs when touching PII or regulated flows.

Start today

While Day 1 is being prepared:

• Introduction — scope, Python/Rust layout, and reading paths
• Quick Start — install core + run Compose
• Production-grade HTTP service — checklist for operating behind an API
• Poll ingest with the standalone gateway — GitHub poll + GET /health/ingest
• Gateway edge & API gateways — Kong / APISIX patterns, CORS, rate limits
• Built-in pipeline — how Tier 1 retrieval + intelligence stages fit together

---

Series hub: add new days as pages under docs/_tutorials/ and link them from this document, or organize by day-01-*.md, day-02-*.md, as you publish.