Promise Theory
Promise theory (Mark Burgess; formalized with Jan Bergstra) is a method of
analysis for systems of autonomous agents — humans, LLM agents, APIs, and
deterministic automation. It supplies the vocabulary for designing and
diagnosing delegation: promises, acceptances, assessments, breaches, and
renegotiation. This skill is a thin router; load the dense material only when
a row in Load By Need matches your task.
Core model
A promise is an autonomous declaration of intended, but as yet unverified,
behaviour from a promiser to a promisee (body: label Λ, type τ, constraint χ).
Agents are autonomous: no agent can promise another's behaviour. Coordination
emerges from voluntary cooperation — an offer plus an acceptance (a
counter-promise) — never from imposed obligation. Obligations are derived,
non-autonomous impositions (imposition + penalty). Agents keep promises via an
evaluation loop: observe → assess → act, converging on the promised state. The
Downstream Principle: the most downstream party in a promise chain carries the greatest causal responsibility for the outcome.
When to use
Load this skill when any of these triggers matches:
- Modeling delegation between humans and agents — decide who may promise what to whom, and who accepts, in a human + AI workforce.
- Designing capability manifests or agent contracts — declare capabilities and intent with acceptance criteria, verification, and withdrawal semantics.
- Diagnosing coordination failures — explain unkept promises, refused acceptances, or missing assessments in multi-agent work.
- Calibrating trust and verification — decide how much to verify an agent, at what rate, and at what cost.
- Designing self-healing or convergent infrastructure — evaluation loops that observe, assess, and act toward a desired state.
- Converting obligation-based designs to promise-based ones — replace push commands and mandates with voluntary offers and acceptance.
When not to use
- When enforceable centralized control is guaranteed — if you can command and verify compliance directly, promise theory's machinery is overhead, not insight.
- For simple single-agent prompting — one model and one prompt, with no delegation graph to model, needs no promise vocabulary.
- For imperative push-based orchestration scripts that need no consent modeling — a cron job or CI pipeline that runs without acceptance semantics is not a promise system.
- For legal contracts — promise theory is not contract law; it models voluntary intent and assessment, not enforceable legal instruments. Draft real contracts with legal counsel.
- When the user needs a specific tool manual — route to the tool's own skill (for example, cli-builder for CLI conventions) instead of framing the tool with promise theory.
Load By Need
| Need | Load |
|---|
| Re-derive a definition or the formal model (promise, imposition, obligation, bindings, trust, Downstream Principle) | references/foundations.md |
| Learn from CFEngine, IaC, or distributed-systems practice before designing convergent infrastructure | references/applications-infrastructure.md |
| Design coordination between specific humans and agents (manifests, acceptance handshakes, oversight, authority) | references/agent-coordination.md |
| Apply a named pattern — promise manifest, acceptance handshake, agent contract, evaluation loop, breach→renegotiation, redundancy, trust calibration | references/patterns.md |
| Decide how much to verify an agent, set a starting trust level, or wire assessment into evals and observability | references/trust-and-verification.md |
| Diagnose a coordination failure, run the breach taxonomy, or check the theory's limitations | references/diagnosis-and-debugging.md |
| Hit an unfamiliar term while applying this skill | references/glossary.md |
Quick Start
Run these commands from the skill directory (promise-theory/); python3 scripts/promise-contract.py --help lists every command and flag.
- Draft a promise manifest. Copy
templates/promise-manifest.yaml.tmpl to a working file (for example promise-manifest.yaml) and fill the placeholders: agent ids and roles, at least one promise per agent (body, type, target), and at least one expectations entry whose about references a declared promise id.
- Lint it. Run
python3 scripts/promise-contract.py lint promise-manifest.yaml. Exit 0 with full expectation coverage means the manifest is valid; exit 1 names the violations to fix (coverage gaps, dangling acceptances, invalid enums) or reports a malformed file as a parse error — never a traceback. Re-run after each fix until clean.
- Add
--json for machine-readable output. Run python3 scripts/promise-contract.py lint promise-manifest.yaml --json to get a single JSON object on stdout (valid, errors, warnings, coverage, bindings) and nothing else.
- Add
--dry-run to confirm no writes. Run python3 scripts/promise-contract.py lint promise-manifest.yaml --dry-run to repeat the same check; lint is read-only, so nothing is written or modified.
Available Scripts
This skill bundles one script; there are no others to discover. Both commands are read-only (--dry-run is accepted everywhere as a no-op guard).
| Script | Purpose | Invocation |
|---|
scripts/promise-contract.py | Validates and renders promise-theory manifest contracts (restricted-YAML or JSON). lint checks a promise manifest against the promise-manifest v1 schema (exit 0 = valid with full expectation coverage; exit 1 = named lint errors or coverage gaps; exit 2 = usage/IO errors) and render prints a promise-graph summary of agents, promises, bindings, and uncovered expectations. Run lint after drafting or every edit of a manifest until it exits clean, and render when you need a human- or machine-readable view of the coordination model you just built. | python3 scripts/promise-contract.py lint promise-manifest.yaml |
Append --json for machine-readable output (a single JSON object on stdout); render --json gives the same treatment to the graph summary.
Related Skills
| Skill | Route when... |
|---|
| agent-evals-and-observability | You need the assessment layer: evals, guardrails, and observability that verify promises are kept (also routed from references/trust-and-verification.md) |
| agent-council | You need multi-agent debate as structured promise exchange and convergence (also routed from references/agent-coordination.md) |
| workflow-architect | You need to design a workflow as a chain of promises (also routed from references/patterns.md) |
| artifact-pyramids | You need to structure promise-keeping evidence as summaries → analysis → evidence dossiers (also routed from references/trust-and-verification.md) |
| agent-skills | You are authoring or editing an Agent Skills-format skill — the format this skill follows |
| cli-builder | You are building or refactoring the bundled CLI — scripts/promise-contract.py follows cli-builder conventions (non-interactive, --json, --dry-run) |
Gotchas
- Provenance honesty. The direct "promise theory + AI agents" literature is thin and recent (Burgess, "Cooperation in Human and Machine Agents," arXiv:2604.10505, 2026). In the references, claims not verified against a primary source carry
[UNVERIFIED], and the promise-theory → LLM-agent synthesis is labeled EXTRAPOLATION. Preserve those markers; they are what keep this skill honest.
- The theory is "semi-formal." The authors themselves use that term: there is a notation, definitions, lemmas, and rules, but no complete axiomatisation or model theory. The famous ≤50% (impositions) vs ≤100% (promises) claim is an informal heuristic, not a derived result. Use the formalism as a reasoning aid, not a proof system.
- Autonomy is a modeling postulate, not an ideology. It does not claim decentralization is morally right or always better; it is chosen because it forces complete documentation of intended behaviour and exposes failure modes.
- Promise-keeping must be stored as data. CFEngine's documented gap: it reported whether a promise was kept right now, but promise-keeping was never stored as data, so the evaluation loop was incomplete. In a hybrid workforce, record assessments as versioned data (a promise ledger) or trust cannot accumulate.
- Verification loads are an attention/energy budget. The rate at which you check (kinetic mistrust) is spent attention; Burgess & Dunbar model it as a bounded budget. Budget verification cost explicitly and start unknown agents at 50-50 rather than assuming trust or distrust.
Prerequisites
- Python 3 with standard library only;
promise-contract.py requires no third-party packages.
- A manifest to lint or render: copy
templates/promise-manifest.yaml.tmpl and fill the placeholders (see Quick Start) before running either command.
Limitations
- The CLI validates declaration structure, enum values, expectation coverage, and dangling acceptances — it cannot judge whether the promised behaviour is sensible, achievable, or actually kept; assessments live in your promise ledger, not in this tool.
- Promise theory is not contract law: nothing the script validates creates an enforceable legal instrument (see When not to use).
- Lint is a static check at a point in time; it does not observe agents or verify runtime promise-keeping.
Exit Conditions
Stop when the delegation is modeled as a promise set, acceptances and assessments are recorded (or their absence explicitly deferred), and every breach has a renegotiation or escalation path. When diagnosing, stop after three non-converging passes and report the evidence instead of re-litigating the same promises.