Status: proposed (16 Oct 2026). Part of the registry design. Survey facts were checked on 16 Oct 2026 (gh repo view for stars and activity, arXiv abstracts and vendor documentation); abstract-only or unverified claims are marked. The reference definition draft_critique_revise 1.0 validates as publishable with MADE 0.10.0 (one warning, discussed under what the generic loop can run) and tools/catalog/walk_ceremonies.py drives its four paths (accepted at once; rejected, revised, accepted; revision cannot settle, escalated; three rejections, escalated) through the installed made-mcp in a disposable store, including the guards that must refuse.
What the generic loop can run today🔗
/ceremony <definition> drives any published definition (workflows.md): the session's model works each host_callback step with the step's own prompt, hands back fields with axlr_step_done, MADE evaluates the guards and the console claims the next step. The envelope a catalog workflow must fit:
| Mechanism | Available in the generic loop | Catalog rule |
|---|
step_status:<step>:COMPLETED | yes | every transition carries it |
output_field:<step>:<field>=<value> | yes (equality on a scalar) | the step derives any judgement into a scalar (accepted=true, decision="approve", severity="high"); a numeric threshold ("score ≥ 4") is computed by the step and handed back as a boolean |
step_repeat_exhausted:<step> with repeat.max_iterations | yes | every loop has a cap and an exit to a terminal or escalation state |
max_transitions / max_bounces | required by MADE on any cyclic definition | every catalog definition declares both |
roles | a step is claimable only by a declared role whose allowed_actions name it and its triggers | every catalog definition declares one role, AXLR, with every step and trigger; without it MADE refuses the first claim ("no workflow role can execute step") |
repeat.max_iterations | counted per entry into the state: a step re-entered after a transition starts its budget again (walked on 16 Oct 2026) | overall rounds are bounded by a field the model derives (the critic counts its rejections) and by max_transitions |
type: human guard (manual_approval) | the loop waits for an approval the console cannot give | not used; a decision of the person is a step whose prompt says to put the question to the person in the conversation and hand back decision. MADE warns that a cycle has "no human-controlled state"; the warning is accepted and shown on the workflow's page |
num_agents > 1 | the console works the step once (no code reads num_agents) | independent voices (council members, debaters, blind screeners, Delphi panellists) are sequential steps with see_prior: false, one per voice, and the collecting step reads them all; every voice is the session's model, since the generic loop cannot pick a model per step, so a "council" is one model in several fresh contexts until a per-step model exists. True fan-out waits for the x-axlr.fan_out extension |
check commands (check_command exits zero) | only in the console's pinned workflows | the step runs the command through local_exec and reports passed; the person approves the command as usual |
| fresh-context review | only in /incident | approximated by a step whose prompt forbids reading the prior transcript (see_prior: false) |
| budget | max_session_usd and the 75K-token session cap apply to the whole session | no per-ceremony guard; the cap and max_transitions bound the work |
The registry design lists the console extension (x-axlr step annotations for approval cards, check commands and fan-out) that would promote the "approximated" rows to real guards. Until then every catalog definition is written inside this envelope and its page says which judgements the model makes for itself.
Survey: orchestration patterns worth packaging🔗
Canonical taxonomies and frameworks🔗
| Source | Patterns | Where a person decides | Evidence of value |
|---|
| Anthropic, Building effective agents (Dec 2024) | prompt chaining, routing, parallelization (sectioning, voting), orchestrator-workers, evaluator-optimizer; autonomous agents last | checkpoints and blockers; max-iteration stops | guidance: the simplest workflow that works |
| Anthropic multi-agent research system | lead plans and saves the plan, parallel subagents, citation agent | none built in; checkpoints | +90.2 % over single Opus 4 on an internal research eval; ~15× the tokens of chat; a poor fit for most coding |
Microsoft Agent Framework (AutoGen's successor; microsoft/autogen archived) | sequential, concurrent, handoff, group chat, magentic (task and progress ledgers, stall counter, replan with caps) | plan review (RequirePlanSignoff), tool approval | Magentic-One competitive on GAIA, AssistantBench, WebArena |
| LangGraph | the Anthropic five plus plan-and-execute, reflection, Reflexion, LATS, supervisor and hierarchical teams | interrupts and checkpoints | 43k stars; the supervisor package now steers to tool-based supervisors |
| CrewAI | sequential and hierarchical processes, flows | human_input on a task | 60k stars |
| OpenAI Agents SDK | handoffs as transfer_to_x tools, agents as tools, guardrails on first and last agent | guardrails | 30k stars |
| Google ADK | sequential, parallel, loop workflow agents; ADK 2.0 moves to graph-based dynamic workflows | | 22k stars |
| AWS Strands | graph, swarm, workflow, agents-as-tools (repo renamed harness-sdk) | | 9k stars |
| Claude Agent SDK subagents | fresh context per subagent, summary back, parallel up to 20, depth 3 | only the parent asks the person | |
| MetaGPT, ChatDev, CAMEL | SOPs as role assembly lines; chat chains with "communicative dehallucination"; two-agent role play | none | 71k / 35k / 18k stars; MetaGPT's last push Jan 2026 |
| spec-kit (GitHub), Kiro specs | constitution → specify → plan → tasks → implement, with clarify, checklist and analyze; requirements/design/tasks documents each reviewed | per phase or per document | spec-kit 141k stars, the strongest adoption signal of the survey |
| Aider architect/editor | a reasoning model describes the change, an editor model writes the edits | interactive | 85 % on aider's edit benchmark |
Research patterns🔗
| Pattern | Structure | Evidence |
|---|
| Reflexion | act → evaluate → written reflection → retry, capped | 91 % pass@1 HumanEval vs 80 % (abstract) |
| Self-Refine | one model produces, critiques, revises | ~+20 points average across 7 tasks (abstract) |
| Tree of Thoughts, LATS | branch, evaluate, prune, backtrack; tree search with reflection | Game of 24: 74 % vs 4 %; LATS 92.7 % HumanEval (abstracts) |
| Multi-agent debate, LLM Council | N proposers, rounds of critique, a chair synthesises; Karpathy's llm-council ranks answers anonymously | better factuality (abstract); council 25k stars |
| Mixture of Agents | layered proposers then an aggregator | 65.1 % AlpacaEval 2.0 vs 57.5 % GPT-4o |
| STORM / Co-STORM | perspectives → simulated interviews → outline → article; Co-STORM keeps a person steering | +25 points organisation vs RAG baseline; 32k stars |
| AI Scientist, Agent Laboratory, Co-Scientist | idea → code → experiments → paper → simulated review; Agent Laboratory adds human feedback at each stage (84 % lower cost than fully autonomous); Co-Scientist runs a hypothesis tournament with debate | in-vitro confirmations for Co-Scientist |
| AgentReview | reviewers, area chair, authors | reviewer bias shifted decisions ~37 % |
| Habermas Machine (Science 2024) | statements → critiques → regenerated statements maximising group approval | preferred to human mediators by >5,000 participants (secondary sources; check the DOI 10.1126/science.adq2852) |
Three regularities that shape the catalog🔗
- Multi-agent value is strongest in breadth-first work (research, review, consultation), weakest and most expensive in coding. Research workflows fan out; software workflows stay sequential with a fresh reviewer.
- Every mature framework converges on the same three controls: an explicit plan a person signs off, loops with hard caps and stall detection, and an independent evaluator with fresh context. MADE already has the first two (
manual_approval, repeat, max_transitions); the catalog models the third as a see_prior: false step.
- Frameworks churn (AutoGen archived, ADK rewrote its workflows, Strands renamed): a declarative YAML catalog is a stable layer above them, and the pattern names in the index (
pattern: in the manifest) use MADE's own catalog where it has one: roundtable_fixed_order, broadcast_collect, group_chat, maker_checker, handoff, magentic, advisor.
Catalog🔗
Tier A runs in the generic loop today; tier B runs today with the person's decision and parallel voices approximated as described above (its page says so); tier C waits for x-axlr guards (approval card, check command, fan-out) to be worth publishing. Rounds are always capped; the cap is in the definition.
Writing and editing🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
draft_critique_revise | polished text from a brief with named criteria | draft → critique → revise ⟲ → deliver / escalate | ≤ 3 critique rounds, ≤ 2 revision attempts each | maker_checker, Self-Refine | A (reference, validated and walked) |
editorial_board | one draft read by fact, style, legal and audience editors, merged by a chief | draft → fact → style → legal → audience → merge → chief verdict | one pass; chief may return once | roundtable_fixed_order | A |
translation_backtranslation | translation checked by a blind back-translation | translate → back-translate (see_prior: false) → diff → fix ⟲ → sign-off | fix ≤ 3 | maker_checker | A |
writers_room | story development | premise → pitches (4 sequential voices) → selection → beat sheet → scene drafts → table read → revise | revise ≤ 2 | broadcast_collect | B |
Research and science🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
literature_review_storm | cited survey of a topic | perspectives → interviews (one step per perspective) → outline → person's outline decision → draft → citation check → revise ⟲ | revise ≤ 2 | STORM, broadcast_collect | B |
systematic_review | screen studies against inclusion criteria, two blind screeners | protocol → search → screen A → screen B (see_prior: false) → conflicts → extraction → bias appraisal → synthesis | conflicts to the person | voting parallelization | B |
reproducibility_check | re-run a paper's claim | claims → environment → run ⟲ → compare → discrepancy report | run ≤ 3 | Reflexion | C (check commands) |
peer_review_simulation | pre-submission critique | three reviews (sequential, see_prior: false) → meta-review → rebuttal → revised verdict ⟲ | until scores stop moving, ≤ 2 | AgentReview, debate | B |
hypothesis_tournament | rank candidate hypotheses | generate → pairwise debate ⟲ → ranking → person's pick → experiment design | debate ≤ 3 rounds | Co-Scientist, group_chat | B |
Software🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
spec_driven_delivery | idea to feature through reviewed specs | specify → clarify ⟲ → plan → tasks → implement (hands off to /plan) → verify → converge | clarify until no ambiguity, ≤ 3 | spec-kit, Kiro, orchestrator-workers | C (approval cards; /plan handoff) |
security_review_council | threat model, injection, authorization, secrets and dependency lenses on a diff | scope → five lenses (sequential, see_prior: false) → consolidate → rebuttal → verdict → remediation plan | rebuttal once | broadcast_collect, council | B |
dependency_upgrade | upgrade one dependency safely | inventory → changelog digest → branch → fix breakages ⟲ → tests → PR | fix ≤ 5 | Reflexion | C (check commands, PR flow of /improve) |
code_migration | migrate in slices with parity checks | survey → strategy (person decides) → slice plan → migrate slice ⟲ → parity → cutover | per slice | plan-and-execute, waves | C |
documentation_audit | find and fix documentation that diverges from code | inventory → claims → verify claims → fix → fresh review | verify per claim | evaluator-optimizer | B |
test_first_loop | TDD on one behaviour | write failing test → red → implement ⟲ → green → refactor | implement ≤ 6 | TDD, Reflexion | C (check commands; /plan tasks already do red/green) |
release_readiness | go or no-go | evidence → checklist → risk review → rollback plan → decision | none | prompt chaining with a gate | B |
Decision-making🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
red_team_review | attack a plan before committing to it | summary → attacks (3 voices, see_prior: false) → defence → severity rating → mitigations ⟲ → sign-off | until no new high finding, ≤ 3 | debate, maker_checker | A |
delphi_estimate | expert-style forecast | question → five estimates (see_prior: false) → anonymised feedback → revise ⟲ → aggregate | until spread under the brief's bound, ≤ 3 | Delphi method (methodology source unverified), Mixture of Agents | B |
pre_mortem | how the project fails | assume failure → five failure stories → cluster → rank → mitigations | none | broadcast_collect | A |
decision_memo_council | options memo with a vote and a dissent record | frame → options → analysis per option → five votes (see_prior: false, the same model five times) → dissent → decision | vote once | LLM Council, group_chat | A |
advisor_checked_task | a task the executor must clear with an advisor before and after | brief → advice → work ⟲ → advice → deliver | advice loops until no blocking advice, ≤ 3 | MADE advisor | A |
Business and administration🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
rfp_response | answer a request for proposals | requirements → compliance matrix → draft per section ⟲ → review → price check → sign-off | per section until the review passes | orchestrator-workers | B |
procurement_comparison | compare vendors against the person's criteria | criteria (person decides) → gather → score per vendor → sensitivity check → recommendation | none | parallelization | B |
meeting_to_actions | notes to owned, confirmed actions | extract → owners → confirm with the person → publish | until every item has an owner | prompt chaining | A |
okr_review | quarterly objectives review | collect evidence → score key results → explain gaps → propose next quarter → person decides | none | prompt chaining | A |
incident_comms | customer-facing communications during an incident | facts → audiences → drafts → legal read → approval → publish | legal read ≤ 2 | maker_checker | B |
Policy and civics🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
policy_brief_stakeholders | a brief tested against simulated stakeholders | evidence → draft → stakeholder simulation (4–6 personas, sequential) → revise ⟲ → sign-off | until objections are addressed, ≤ 3 | Generative Agents, Habermas | B |
legislative_impact | impact analysis of a bill | summary → affected groups → cost and benefit per group → legal consistency → options → brief | none | prompt chaining, broadcast_collect | A |
consultation_synthesis | synthesis of public comments | cluster → positions → common ground → statement → critique rounds ⟲ → final statement | until approval rises no further, ≤ 3 | Habermas Machine | B |
deliberative_minipublic | a simulated citizens' assembly on one question | briefing → small-group rounds (sequential) → plenary → recommendations → minority report | rounds ≤ 3 | group_chat | B |
Education and creative🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
lesson_design_with_simulation | lesson plan tried on simulated students | goals → plan → three students → misconception log → revise ⟲ → exam → validation | until the students pass, ≤ 3 | Self-Refine plus simulation | B |
socratic_tutoring | tutoring that asks, never tells, until the learner explains it back | diagnose → question ⟲ → learner explanation → check → next concept | per concept, ≤ 6 turns | evaluator-optimizer | A |
exam_generation | an exam with validated answers | blueprint → items → blind solve (see_prior: false) → disagreement fix ⟲ → key | fix ≤ 2 | maker_checker | A |
game_design_loop | mechanics tested by simulated playtests | pitch → rules → three playtests → findings → revise ⟲ → rulebook | revise ≤ 3 | broadcast_collect, maker_checker | B |
Data and personal🔗
| Name | Goal | States | Loop | Pattern | Tier |
|---|
analysis_plan_execute_audit | an analysis with a blind audit | plan → execute → audit (see_prior: false) → fix ⟲ → report | fix ≤ 2 | plan-and-execute | B |
data_cleaning_gates | cleaning with validation gates | profile → rules → apply → validate ⟲ → sign-off | until validation passes, ≤ 3 | evaluator-optimizer | C (check commands) |
weekly_review | a personal weekly review | collect → review → prioritise → plan → commit | none | prompt chaining | A |
First batch🔗
In order of value over cost of console work: draft_critique_revise (done), decision_memo_council, red_team_review, pre_mortem, editorial_board, meeting_to_actions, exam_generation, socratic_tutoring, legislative_impact, advisor_checked_task — all tier A — then the tier B research set (literature_review_storm, security_review_council, peer_review_simulation). spec_driven_delivery is the most-adopted pattern in the survey but needs the approval card and a handoff to /plan, so it is the first tier-C candidate once the x-axlr extension exists.
Sources🔗
Anthropic: Building effective agents, multi-agent research system. Microsoft: Agent Framework orchestrations, Magentic-One. LangGraph workflows, CrewAI processes, OpenAI Agents SDK handoffs, Google ADK workflow agents, Strands multi-agent patterns, Claude Code subagents, spec-kit, Kiro specs, Aider architect mode. Papers: Reflexion, Self-Refine, Tree of Thoughts, LATS, multi-agent debate, Mixture of Agents, STORM, Co-STORM, AI Scientist, Agent Laboratory, Co-Scientist, Generative Agents, AgentReview, Habermas Machine, MetaGPT, ChatDev, CAMEL, SWE-agent, OpenHands, llm-council.
