Agent workflows

AXLR runs ordinary work directly. A workflow starts only when you ask for one: a work-mode command, /ceremony with a definition, or a self-repair or self-improvement the agent requests with evidence the console validates. The console drives the instance and the model does each step's work with AXLR tools; the model never drives a console workflow through direct made_* calls. MADE records every accepted step completion and transition in a hash-chained journal, so the result is an instance whose history can be checked, not a state that changed by itself. MADE does not verify the work: that a step did real work is the console's evidence (command exits, revisions, files), which MADE records.

MADE

MADE is Underpass's open-source coordination engine (Apache-2.0, github.com/underpass-ai/made, pre-1.0). It runs no agents and, in its local form, calls no model: the console does each step's work, while MADE decides which step may be claimed next and enables a transition only when its guards hold. Its journal verifies its own integrity and order, not the truth of the work, and a person's approval at a human guard is a separate recorded fact that an agent's claim cannot replace. AXLR uses MADE locally, as one MCP process (made-mcp) over one SQLite file; MADE also runs as a gRPC service.

Built in

Six commands start the console's pinned definitions. Selecting /debug, /delivery or /incident does not start an instance; the next prompt does. /plan opens the plans panel, and /repair and /improve run only in a clone that axlr-tui --repair or --improve prepared. AXLR checks the published definition's semantic digest, starts the instance and claims its first step.

Command	Definition	Sequence and acceptance
/debug	axlr_debug 2.0	Reproduce with a nonzero command exit → diagnose → repair until the same command exits zero → integrate
/delivery	axlr_delivery 2.0	Brief with criteria, scope and baseline command → build until that command exits zero → integrate
/incident	axlr_incident 1.0	Triage → evidence timeline → analysis and actions → draft and fresh-context review → the person's approval → postmortem publication
/plan	axlr_plan, axlr_task, axlr_sync 1.0	Decompose a brief into verified atomic tasks → the person's approval → each task in a fresh worker (red for test-first tasks, green, hand-back) → a sync after each wave
/repair	axlr_repair 1.0	Reproduce → diagnose → repair in a fresh clone → the console commits, pushes and opens the pull request → watches its checks → merge decision (automatic with repair.auto_merge, otherwise yours) → squash merge
/improve	axlr_improve 1.0	Brief with a check that fails in a fresh clone → build until it passes → the console opens the pull request → watches its checks → the person's merge decision → squash merge

• Check commands. A proposed check is {program, args} with no implicit shell. The first command, or a changed reproduction command, needs your approval even in full autonomy; once accepted, repair and build always rerun the saved command. Each run has a five-minute timeout.
• Bounded loops. Debug reproduction and repair, and delivery build, allow at most three iterations per phase. Exhaustion leads to BLOCKED; successful integration to COMPLETED. Both return the session to normal mode.
• What integration proves. In /debug and /delivery, integration records the report, git rev-parse HEAD and the dirty files; it does not commit, merge or publish. A zero exit proves only what that command checks.
• Self-repair and self-improvement. axlr-tui --repair "<brief>" and axlr-tui --improve "<brief or #issue>" let the console change the configured repository (repair.repository, AXLR's own by default) through a pull request it opens with gh; an improvement always leaves the merge to you.

The console's workflow guide has the full behaviour, and the MADE runbook prepares the identity and publishes the pinned definitions.

Any published definition

/ceremony lists the published definitions; /ceremony <definition> [version] [brief] runs one through the generic loop. The console claims each step MADE lists as claimable, hands the model the step's own prompt from the definition, completes the step with exactly the fields the model hands back through axlr_step_done, applies the transition MADE enables and claims the next step. At a terminal state it writes the outcome to KMP and returns the session to normal mode. A state with no enabled transition and no claimable step, such as a MADE human guard, waits for something outside the console.

/ceremony publish <file.yaml> validates and publishes a definition you wrote; a relative path is read from the workspace. A published name and version is immutable: change the version when the content changes, and running instances stay on the version they started with. The model can design definitions; it never publishes one.

/ceremony publish draft_critique_revise-1.0.yaml
/ceremony draft_critique_revise a 300-word release note; criteria: accurate, plain, names every breaking change

The catalog

The catalog packages orchestration patterns that recur across agent frameworks and research (maker-checker, councils, red teams, Delphi rounds, STORM-style research, spec-driven delivery) as MADE definitions the generic loop can run. It is a proposal of 16 October 2026: one reference definition exists, and the console's /ceremony install command and the catalog pages per workflow are designed but not built yet. Until then, a catalog definition is published with /ceremony publish.

The reference definition: draft_critique_revise 1.0

Write a draft from a brief, critique it against the brief's named criteria, revise; at most three critique rounds (counted by the critic and bounded by MADE's max_bounces) and two revision attempts per round (MADE's step repeat). Deliver when the critique accepts it, escalate to the person when the rounds run out. It validates as publishable with MADE 0.10.0, with one accepted warning (its loop has no human-controlled state), and its four paths (accepted at once; rejected, revised, accepted; a revision that cannot settle; three rejections) were walked through a real made-mcp in a disposable store. Read the definition.

stateDiagram-v2
    direction LR
    [*] --> DRAFT
    DRAFT --> CRITIQUE: drafted
    CRITIQUE --> DELIVER: accepted
    CRITIQUE --> REVISE: revision needed
    CRITIQUE --> ESCALATE: critique exhausted
    REVISE --> CRITIQUE: revised
    REVISE --> ESCALATE: rounds exhausted
    DELIVER --> COMPLETED: delivered
    ESCALATE --> BLOCKED: escalated
    COMPLETED --> [*]
    BLOCKED --> [*]

What the generic loop can run today

Mechanism	Today	Rule for a catalog definition
Automated guards on step status (step_status:<step>:COMPLETED) and on a top-level output field (output_field:<step>:<field>=<json>, exact equality)	yes	every judgement is derived into a scalar (accepted=true, decision="approve"); a numeric threshold is computed by the step and handed back as a boolean
Loops: repeat.max_iterations on a step or state, max_transitions for the whole instance, max_bounces per edge	yes	every loop has a cap and a step_repeat_exhausted or state_repeat_exhausted exit to a terminal or escalation state
A person's decision	approximated	a step whose prompt puts the question to the person in the conversation and hands back decision, which is the model's output, not a recorded human approval. The generic loop cannot approve a MADE human guard (type: human, check: manual_approval); the built-in workflows approve theirs through a separate approver identity
Independent voices (councils, debaters, blind screeners)	approximated	sequential steps, one per voice, whose prompt tells the model not to rely on earlier voices (marked see_prior: false). MADE reads that flag only in its own council handler and the generic loop runs every voice in the same session, so independence rests on the prompt alone
Check commands	only in the pinned built-in workflows	the step runs the command through local_exec and reports passed
Budget	per session	max_session_usd, the per-request context cap (prompt_tokens, 75,000 by default) and max_transitions bound the work

Tiers

A runs in the generic loop today. B runs today with the person's decision and parallel voices approximated as above; its page will say so. C waits for console features (an approval card in the generic loop, check commands, fan-out) before it is worth publishing. Rounds are always capped; the cap is in the definition.

In the tables below, a "council" is one model given several prompts, not a MADE council of registered agents, and "blind" means a prompt that excludes earlier voices. The snake_case pattern names are MADE's own (maker_checker, broadcast_collect, group_chat, handoff, magentic, advisor) or its roundtable_fixed_order preset; catalog definitions approximate them with ordinary sequential states and steps. MADE's advisor kind needs MADE 0.10.0.

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 the 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.

The workflow catalog design has the survey behind these choices, with its sources.