agentcy

Consolidated monorepo for the Agentcy CLI suite. Member runtimes transform a durable Brand / Voice / Visual / Content / Outcomes foundation into explicit protocol artifacts.

Read these first:

• docs/capability-model.md — foundation, runtimes, extensions
• docs/principal-patterns.md — small-core operating model
• docs/design-md-fidelity.md — DESIGN.md evidence-first visual contract

A portable brand kit lives at brands/<brand>/{BRAND.md, DESIGN.md, brand.yml, assets/}. BRAND.md is the behavioral contract; DESIGN.md follows Google's public design.md shape under design_system / tokens / components, with Agentcy execution rules under agentcy.

Members

Naming contract

Each protocol artifact's writer.repo and writer.module carry the same agentcy-* name. Python imports match with underscores (agentcy_voice, agentcy_briefs, agentcy_forecast, agentcy_measure).

Capabilities and extensions

See docs/capability-model.md for the canonical model.

Current extension families:

• Studio generation: social.post, blog.post, outreach.touch, respond.reply, lab card/cover/render
• Forecasting: audience/social reaction forecasts and repo-local run_eval sidecars
• Measurement: adapt, calibrate, study
• Agent instructions: one nested skills/agentcy skill with reference files

Natural next Studio modes are og.cover, social.card, carousel, short.video, ugc.ad, longform.script, and campaign.pack.

Setup

# Python workspace + repo-local dev tools
uv sync --group dev

# Studio runtime (under studio/)
cd studio && pnpm install

# Forecast full simulation runtime, isolated on Python 3.11
make install-forecast-simulation

Best fit

Agentcy is best when you need a protocol-first workflow stack rather than a single embedded SDK:

• brand kit → brief → Studio artifacts by default, with Voice/Forecast/Measure as opt-in extensions
• multi-tenant brand execution from agent-facing BRAND.md contracts
• human-and-agent collaboration over stable JSON/file handoffs
• resumable operator workflows with explicit artifacts and lineage

It is not yet the best fit for:

• non-technical one-click onboarding
• teams that only want a single import-and-go library

Install profiles

The suite is consumable in layers:

# Base Python suite: root CLI + protocols + voice + briefs + forecast base CLI + measure
uv sync --group dev
# or: make install-python-suite

# Full Forecast simulation runtime (isolated Python 3.11 env)
make install-forecast-simulation

# Studio runtime (Node, under studio/)
cd studio && pnpm install
# or: make install-studio

# Full local operator stack
make install-full-operator

Published package contours:

• agentcy-protocols — closest thing to a drop-in library layer
• agentcy-* member CLIs — stage-owned workflow tools
• agentcy — umbrella dispatcher and pipeline orchestrator

The umbrella package is therefore best understood as an operator CLI suite, not a single drop-in SDK.

Useful discovery commands:

agentcy catalog --json
agentcy quickstart --profile full-operator --json
agentcy doctor --json
agentcy member briefs --json plan list

Test it

make doctor
make check
make lint

Direct test commands

# The protocol seam tests shell into studio, so install the runtime once in a clean checkout.
cd studio && pnpm install

uv run pytest tests briefs/tests forecast/tests measure/tests voice/tests protocols/tests -q
cd studio && pnpm check

Pipeline commands

Root pipeline helpers

# Essential bundle: cheap deterministic output under artifacts/pipelines/<pipeline_id>/
uv run agentcy pipeline run \
  --pipeline-id givecare-launch-01 \
  --brand givecare \
  --brief "Before fall gets busy, make caregiving feel lighter" \
  --mode preview \
  --output-dir artifacts/pipelines \
  --json

# Default path:
# brief.v1 -> Studio draft/render -> inspectable artifacts
# It does not run Voice, Forecast, Measure, providers, or publish unless requested.

# Heavier opt-in path
uv run agentcy --provider claude-cli --model sonnet pipeline run \
  --pipeline-id givecare-launch-forecast-01 \
  --brand givecare \
  --brief "Before fall gets busy, make caregiving feel lighter" \
  --persona scientist \
  --persona-eval \
  --with-forecast \
  --files docs/launch-memo.md \
  --publish \
  --smoke \
  --output-dir artifacts/pipelines \
  --json

# After Studio publish + Measure adapt happen, backfill the bundle with canonical later-stage artifacts
uv run agentcy pipeline update \
  --manifest artifacts/pipelines/<pipeline_id>/manifest.json \
  --run-result /tmp/run_result.json \
  --performance /tmp/performance.json \
  --json

# Re-open the manifest later and run Measure study once performance exists
uv run agentcy pipeline study \
  --manifest artifacts/pipelines/<pipeline_id>/manifest.json \
  --json

Dispatcher commands also accept root-level LLM overrides that are forwarded to members which honor them. The root pipeline also supports --pipeline-id so stable bundles can land at paths like artifacts/pipelines/givecare-launch-01/:

uv run agentcy --provider claude-cli --model haiku forecast run --files docs/memo.md --requirement "Predict reaction" --smoke --json

Live pipeline

uv run agentcy-voice --json export scientist --to voice-pack.v1 > /tmp/voice_pack.json
uv run agentcy-briefs plan run "Before fall gets busy, make caregiving feel lighter" \
  --brand givecare \
  --voice-pack-input /tmp/voice_pack.json \
  --brief-v1-output /tmp/brief.json \
  -f json > /tmp/brief_plan.json
forecast/.venv-simulation/bin/agentcy-forecast run --files docs/ --brief /tmp/brief.json --json > /tmp/forecast.json
uv run agentcy studio run social.post --brand givecare --brief-file /tmp/brief.json --json > /tmp/run_result.json
uv run agentcy-measure adapt --run-result /tmp/run_result.json --sidecar sidecar.json --output /tmp/performance.json --json > /tmp/performance.stdout.json
uv run agentcy-measure calibrate --forecast /tmp/forecast.json --performance /tmp/performance.json --json > /tmp/calibration.json

Runtime constraints

• agentcy-forecast full simulation requires make install-forecast-simulation, which creates forecast/.venv-simulation on Python 3.11
• agentcy-forecast run --smoke skips ontology/graph/profiles and the OASIS subprocess, then emits deterministic run artifacts for plumbing checks
• Studio runtime under studio needs pnpm install
• make pipeline-fixtures is the fixture-backed smoke path when you only want to validate downstream protocol plumbing

Current status

• Root dispatcher: healthy, probes member reachability, exposes pipeline run / pipeline update / pipeline study, defaults to Briefs mock provider plus deterministic Studio artifacts, forwards root-level --provider / --model overrides when explicitly provided, and exposes agentcy member <member> --json ... as a normalized wrapper over member-local JSON differences
• Voice: structured eval tiers and saved eval-report review flow now ship in the CLI
• Briefs: healthy for package/CLI + planning surfaces, supports provider/model forwarding, and exposes both --json preference and --json-envelope normalized success envelopes
• Forecast: base CLI healthy; completed runs emit a repo-local run_eval sidecar alongside canonical forecast export; full simulation still requires the isolated Python 3.11 env
• Studio: social rendering is deterministic SVG/resvg with no Gemini/image-provider dependency in the default workflow; lab card/cover/render are the first visible extension surfaces
• Measure: supports standardized --json envelopes and a study command that ingests optional forecast/voice eval sidecars