Contracts before code. Tests as law. Agents that can't cheat.
Pact is a multi-agent software engineering framework where the architecture is decided before a single line of implementation is written. Tasks are decomposed into components, each component gets a typed interface contract, and each contract gets executable tests. Only then do agents implement -- independently, in parallel, even competitively -- with no way to ship code that doesn't honor its contract. Generates Python, TypeScript, or JavaScript.
The insight: LLMs are unreliable reviewers but tests are perfectly reliable judges. So make the tests first, make them mechanical, and let agents iterate until they pass. No advisory coordination. No "looks good to me." Pass or fail.
Pact is for projects where getting the boundaries right matters more than getting the code written fast. If a single Claude or Codex session can build your feature in one pass, just do that -- Pact's decomposition, contracts, and multi-agent coordination would be pure overhead.
Use Pact when:
- The task has multiple interacting components with non-obvious boundaries
- You need provable correctness at interfaces -- not "it seems to work" but "it passes 200 contract tests"
- The system will be maintained by agents who need contracts to understand what each piece does
- You want competitive or parallel implementation where multiple agents race on the same component
- The codebase is large enough that no single context window can hold it all
Don't use Pact when:
- A single agent can build the whole thing in one shot
- The task is a bug fix, refactor, or small feature
- You'd spend more time on contracts than on the code itself
Tested on 5 ICPC World Finals competitive programming problems (212 test cases total) using Claude Opus 4.6.
| Condition | Pass Rate | Cost |
|---|---|---|
| Claude Code single-shot | 167/212 (79%) | $0.60 |
| Claude Code iterative (5 attempts) | 196/212 (92%) | $1.26 |
| Pact (solo, noshape) | 212/212 (100%) | ~$13 |
Pact's contract-first pipeline solves problems that iterative prompting cannot. On Trailing Digits (2020 World Finals), Claude Code scores 31/47 even with 5 retry iterations and full test feedback -- the naive algorithm times out on large inputs. Pact's interview and decomposition phases force upfront mathematical analysis, producing the correct O(log n) approach on the first implementation attempt.
Full results: icpc_official/RESULTS.md in the benchmark directory.
Pact treats contracts as source of truth and implementations as disposable artifacts. The code is cattle, not pets.
When a module fails in production, the response isn't "debug the implementation." It's: add a test that reproduces the failure to the contract, flush the implementation, and let an agent rebuild it. The contract got stricter. The next implementation can't have that bug. Over time, contracts accumulate the scar tissue of every production incident -- they become the real engineering artifact.
This inverts the traditional relationship between code and tests. Code is cheap (agents generate it in minutes). Contracts are expensive (they encode hard-won understanding of what the system actually needs to do). Pact makes that inversion explicit: you spend your time on contracts, agents spend their time on code.
The practical upside: when someone asks "who's debugging this at 3am?" -- agents are. The Sentinel watches production logs, detects errors, attributes them to the right component via embedded PACT log keys, spawns a knowledge-flashed fixer agent loaded with the full contract/test context, adds a reproducer test, rebuilds the module, and verifies all tests pass. The contract ensures they can't introduce regressions. The human reviews the contract change in the morning, not the code.
git clone https://github.com/jmcentire/pact.git
cd pact
make
source .venv/bin/activateThat's it. Now try:
pact init my-project
# Edit my-project/task.md with your task
# Edit my-project/sops.md with your standards
pact --helpTask
|
v
Interview -----> Shape (opt) -----> Decompose -----> Contract -----> Test
| | | |
v v v v
Pitch: appetite, Component Tree Interfaces Executable Tests
breadboard, risks |
v
Implement (parallel, competitive)
|
v
Integrate (glue + parent tests)
|
v
Polish (Goodhart tests + regression check)
|
v
Diagnose (on failure)
Ten phases, all mechanical gates:
- Interview -- Establish processing register, then identify risks, ambiguities, ask clarifying questions
- Shape -- (Optional) Produce a Shape Up pitch: appetite, breadboard, rabbit holes, no-gos
- Decompose -- Task into 2-7 component tree, guided by shaping context if present
- Contract -- Each component gets a typed interface contract
- Test -- Each contract gets executable tests plus hidden Goodhart tests (adversarial acceptance criteria the implementation agent never sees)
- Validate -- Mechanical gate: refs resolve, no cycles, tests parse
- Implement -- Each component built independently by a code agent
- Integrate -- Parent components composed via glue code
- Polish -- Cross-component regression check + Goodhart test evaluation with graduated-disclosure remediation
- Diagnose -- On failure: I/O tracing, root cause, recovery
Pact monitors its own coordination health — detecting the specific failure modes of agentic pipelines before they consume the budget.
| Metric | What It Detects |
|---|---|
| Output/planning ratio | Spending $50 on planning and shipping nothing |
| Rejection rate | Agents optimizing for each other's approval, not outcomes |
| Budget velocity | Coordination cost exceeding execution value |
| Phase balance | Any single phase consuming disproportionate budget |
| Cascade detection | One component's failure propagating through the tree |
| Register drift | Agent departing from established processing mode mid-task |
When critical conditions fire, Pact pauses and proposes remedies via FIFO — the user decides whether to apply them. The system never silently modifies its own configuration.
pact health my-project # View health metrics
pact signal my-project --directive \
'{"type":"apply_remedy","remedy":"max_plan_revisions","value":1}'Per-project thresholds in pact.yaml:
health_thresholds:
output_planning_ratio_warning: 0.3
rejection_rate_critical: 0.9
cascade_critical: 10| Lever | Config Key | Effect |
|---|---|---|
| Parallel Components | parallel_components: true |
Independent components implement concurrently |
| Competitive Implementations | competitive_implementations: true |
N agents implement the SAME component; best wins |
Either, neither, or both. Defaults: both off (sequential, single-attempt).
Set plan_only: true to stop after contracts and tests are generated. Then target specific components:
pact components my-project # See what was decomposed
pact build my-project sync_tracker # Build one component
pact build my-project sync_tracker --competitive --agents 3| Command | Purpose |
|---|---|
pact init <project> |
Scaffold a new project |
pact run <project> |
Run the pipeline |
pact daemon <project> |
Event-driven mode (recommended) |
pact status <project> [component] |
Show project or component status |
pact components <project> |
List components with status |
pact build <project> <id> |
Build/rebuild a specific component |
pact interview <project> |
Run interview phase only |
pact answer <project> |
Answer interview questions |
pact approve <project> |
Approve with defaults |
pact validate <project> |
Re-run contract validation |
pact design <project> |
Regenerate design.md |
pact stop <project> |
Gracefully stop a running daemon |
pact log <project> |
Show audit trail (--tail N, --json) |
pact ping |
Test API connection and show pricing |
pact signal <project> |
Resume a paused daemon |
pact watch <project>... |
Start Sentinel production monitor (Ctrl+C to stop) |
pact report <project> <error> |
Manually report a production error |
pact health <project> |
Show health metrics, findings, and proposed remedies |
pact incidents <project> |
List active/recent incidents |
pact incident <project> <id> |
Show incident details + diagnostic report |
pact audit <project> |
Spec-compliance audit (compare task.md vs implementations) |
pact tasks <project> |
List phase tasks with status |
pact analyze <project> |
Run cross-artifact analysis |
pact checklist <project> |
Generate requirements checklist |
pact test-gen <project> |
Generate tests + security audit for any codebase |
pact adopt <project> |
Adopt existing codebase under pact governance |
pact handoff <project> <id> |
Render/validate handoff brief for a component |
pact pricing |
Show model pricing table (--export to override) |
Global (config.yaml at repo root):
model: claude-opus-4-6
default_budget: 10.00
parallel_components: false
competitive_implementations: false
competitive_agents: 2
max_concurrent_agents: 4
plan_only: false
# Override token pricing (per million tokens: [input, output])
model_pricing:
claude-opus-4-6: [5.00, 25.00]
claude-sonnet-4-5-20250929: [3.00, 15.00]
claude-haiku-4-5-20251001: [0.80, 4.00]
# Production monitoring (opt-in)
monitoring_enabled: false
monitoring_auto_remediate: true
monitoring_budget:
per_incident_cap: 5.00
hourly_cap: 10.00
daily_cap: 25.00
weekly_cap: 100.00
monthly_cap: 300.00Per-project (pact.yaml in project directory):
budget: 25.00
parallel_components: true
competitive_implementations: true
competitive_agents: 3
# Shaping (Shape Up methodology)
shaping: true # Enable shaping phase (default: false)
shaping_depth: standard # light | standard | thorough
shaping_rigor: moderate # relaxed | moderate | strict
shaping_budget_pct: 0.15 # Max budget fraction for shaping
# Production monitoring (per-project)
monitoring_log_files:
- "/var/log/myapp/app.log"
- "/var/log/myapp/error.log"
monitoring_process_patterns:
- "myapp-server"
monitoring_webhook_port: 9876
monitoring_error_patterns:
- "ERROR"
- "CRITICAL"
- "Traceback"
# Health thresholds (override defaults)
health_thresholds:
output_planning_ratio_warning: 0.3
rejection_rate_critical: 0.9Project config overrides global. Both are optional.
Route different roles to different providers for cost optimization:
budget: 50.00
role_models:
decomposer: claude-opus-4-6 # Strong reasoning for architecture
contract_author: claude-opus-4-6 # Precision for interfaces
test_author: claude-sonnet-4-5-20250929 # Fast test generation
code_author: gpt-4o # Cost-effective implementation
role_backends:
decomposer: anthropic
contract_author: anthropic
test_author: anthropic
code_author: openai # Mix providers per roleAvailable backends: anthropic, openai, gemini, claude_code, claude_code_team.
Each project is a self-contained directory:
my-project/
task.md # What to build
sops.md # How to build it (standards, stack, preferences)
pact.yaml # Budget and execution config
design.md # Auto-maintained design document
.pact/
state.json # Run lifecycle
audit.jsonl # Full audit trail
decomposition/ # Tree + decisions
contracts/ # Per-component interfaces + goodhart/
implementations/ # Per-component code + contract tests
compositions/ # Integration glue
learnings/ # Accumulated learnings
monitoring/ # Incidents, budget state, diagnostic reports
Beyond contract validation, Pact includes structural checks that catch problems early:
-
North-Star Validation -- After implementation, checks that composed contracts plausibly fulfill the original task. Extracts action verbs from
task.mdand verifies they appear in contract function names/descriptions. Catches the "9 components pass 796 tests but the service can't actually do anything" failure mode. -
Early Decomposition Validation -- Runs after decomposition but before spending LLM calls on contract generation. Checks for orphan nodes, empty descriptions, duplicate descriptions, and task keyword coverage. Saves cost on structurally broken decompositions.
-
Handoff Brief Validation --
pact handoff <project> <component-id> --validatechecks context fence presence, primer ordering, natural format, token budget, and dependency coverage. Useful for debugging agent coordination issues. -
Smoke Test Generation --
pact adoptgenerates mechanical smoke tests from AST analysis -- no LLM required. Extracts all public module-level function signatures (filtering out methods, private functions, and nested functions via indentation check) and produces import + callable check tests intests/smoke/.
Pact ships with a Claude Code custom slash command for crafting optimal task specifications:
/project:craft-task
This interactive command interviews you about your project and generates research-backed task.md, sops.md, and pact.yaml files optimized for Pact's decomposition pipeline.
make dev # Install with LLM backend support
make test # Run full test suite (1886 tests)
make test-quick # Stop on first failure
make clean # Remove venv and cachesRequires Python 3.12+. Core has two dependencies: pydantic and pyyaml. LLM backends require anthropic.
See CLAUDE.md for the full technical reference.
Pact is one of three systems (alongside Emergence and Apprentice) built to test the ideas in Beyond Code: Context, Constraints, and the New Craft of Software. The book covers the coordination, verification, and specification problems that motivated Pact's design.
Baton is a circuit orchestration platform that manages service topologies through a circuit-first design. Pact produces contracted components; Baton wires them into a running topology with mock collapse, A/B routing, health monitoring, and self-healing. Together they cover the full lifecycle: Pact builds the pieces with provable interfaces, Baton runs them in production.
MIT