README

The Laravel AI SDK is how your app talks to an AI. This package is how your app runs a process that involves AI: several steps, decisions, waiting on people, and picking up where it left off.

Picture a real feature: reviewing a contract. Extract the clauses, score the risk, escalate if the risk is high, wait for a manager to sign off, then write the summary. PHP cannot wait until Tuesday for the manager to click approve, so teams build the same pile of glue every time: a pending_reviews table, status columns, jobs, a resume endpoint, retry logic. This package is that pile, done once, on the substrate Laravel already ships: queues, batches, and retries.

A process that can wait. awaitHuman() parks a run in the database for hours or weeks; resume() validates the human's answer and continues. awaitEvent() does the same for webhooks and other systems.
Retry the step that broke, not the whole thing. Every step's result is committed before the next runs; earlier steps keep their results and their token bill stays paid.
Memory between steps. A state bag is saved after every step, so step 5 can read what step 2 produced, even days later on a different server.
A paper trail. Every run and every attempt of every step is a queryable Eloquent record: status, timings, token counts, errors, who approved what and when.

Already know the multi-agent space? This package keeps the vocabulary of the five agent patterns from the Laravel blog's multi-agent walkthrough and makes each one crash-safe: see Five patterns, made durable.

Note

Status: pre-release. The core engine (sequential, conditional, parallel, evaluator, and debate steps; checkpoint/retry; interrupts; events; testing fakes) is implemented and tested. APIs may change before 1.0.

What it looks like

// app/AgentWorkflows/ContractReview.php
class ContractReview extends Workflow
{
    public function build(WorkflowDefinition $workflow): WorkflowDefinition
    {
        return $workflow
            ->step(
                ExtractClausesAgent::class, 
                'Extract the key clauses: {{ contract }}'
            )
            ->step(
                RiskAnalysisAgent::class, 
                'Assess the risk of: {{ output:ExtractClausesAgent }}'
            )
            ->when(
                fn (WorkflowState $state) => $state->output(RiskAnalysisAgent::class)?->structured('riskScore') >= 7,
                then: EscalationAgent::class,
                thenPrompt: 'Draft an escalation memo covering: {{ output:RiskAnalysisAgent }}'
            )
            ->awaitHuman(
                reason: 'Final sign-off required',
                schema: ['approved' => 'required|boolean', 'notes' => 'nullable|string']
            )
            ->step(
                GenerateSummaryAgent::class, 
                'Summarize this review for the record: {{ output:RiskAnalysisAgent }}'
            );
    }
}

$run = ContractReview::start(['contract' => $text]);
// Each step runs as a queued job and is checkpointed as it completes. The run
// parks at the sign-off gate, through deploys, restarts, and weekends.

$run->resume(['approved' => true], by: $request->user());  // the manager answers on Tuesday
$run->retry();                                             // a step failed? re-run that step only;
                                                           // earlier steps keep their results and their token bill

The Quick Start builds a workflow like this end to end.

The step types

Every step type in its simplest form, one acquisition review from end to end:

// app/AgentWorkflows/AcquisitionReview.php
return $workflow
    // any invokable class is a step
    ->step(FetchFilings::class)

    // an agent: one checkpointed agentic turn
    ->step(SummarizeFilingsAgent::class)

    // fan out the analysis, merge when all finish
    ->parallel([
        FinancialAnalysisAgent::class,
        [LegalAnalysisAgent::class, 'Review for legal exposure: {{ output:SummarizeFilingsAgent }}'],
    ])

    // argue the thesis; a judge rules each round
    ->debate(
        ['bull' => BullCaseAgent::class, 'bear' => BearCaseAgent::class],
        judge: VerdictAgent::class,
        as: 'thesis'
    )

    // no consensus? dig deeper, else skip ahead
    ->when(
        fn (WorkflowState $state) => ! $state->get('steps.thesis.satisfied'),
        then: DeepDiveAgent::class
    )

    // revise the memo until it is ready
    ->evaluate(
        DraftMemoAgent::class,
        as: 'memo',
        until: fn (WorkflowState $state) => $state->get('steps.memo.structured.score', 0) >= 8
    )

    // park for a person, for hours or weeks
    ->awaitHuman(reason: 'Partner sign-off required')

    // park until the wire lands
    ->awaitEvent('funds.cleared')

    // close it out
    ->step(RecordInvestment::class);

Each one is documented in Defining Workflows, Human in the Loop, and Agent Debates.

Installation

Requires PHP 8.3+, Laravel 12 or 13, and laravel/ai ^0.10.3:

composer require timmcleod/agent-workflows

The Quick Start takes it from there. Want to see it run first? The demo app is a complete contract-review pipeline (agents, a simulated outage for the retry demo, a risk-based branch, human sign-off, and an agent debate) driven from artisan commands.

Features at a glance

Feature	In one line	Details
`step()`	Chain agents and plain PHP classes; every arrow in the chain is a checkpoint.	Defining Workflows
`when()`	Branch on checkpointed state; the decision is recorded for audit.	Defining Workflows
`parallel()`	Durable fan-out as a queued `Bus::batch`, merged when all branches finish; branches carry their own prompts.	Defining Workflows
`evaluate()`	Loop a step until a predicate passes, checkpointing every iteration, capped.	Defining Workflows
`debate()`	Agents argue in rounds; a judge rules on the transcript after each.	Agent Debates
`awaitHuman()`	Park for sign-off, hours or weeks, with validation schemas and SLA timeouts.	Human in the Loop
`awaitEvent()`	Park until another system delivers a named event (webhooks, payments).	Human in the Loop
Retry from checkpoint	`$run->retry()` re-runs only the failed step; earlier tokens stay paid.	Runs & Observability
Tool-approval bridge	SDK tool approvals park the run; `resume()` replays the decisions.	Human in the Loop
Agent steps	Prompts live on the step, not the agent; a step is one full agentic turn.	Agent Steps
Workflow state	A checkpointed bag with `output()` addressing and typed per-workflow classes.	Workflow State
Audit trail & events	Every attempt an Eloquent row (timings, tokens, errors); lifecycle events for everything.	Runs & Observability
Per-call audit	Each attempt records every provider call inside it: model, tokens, tool calls, invocation id.	Runs & Observability
Testing fakes	`AgentWorkflow::fake()` assertions over really-executing workflows.	Testing
Definition drift	Deploy-changed workflows refuse to resume in-flight runs by default.	Defining Workflows
Operations	A sweeper for dead workers, queue-sizing rules, at-least-once semantics.	Operations

Documentation

Full documentation lives in the documentation index, from the Quick Start through defining workflows, state, human gates, debates, testing, and operations.

What this package is not

Not an arbitrary graph engine. Sequential + conditional + parallel + loop + interrupt covers the overwhelming majority of production workflows. No cycles-with-reducers, no time-travel debugging.
Not a free-form group-chat framework. Structured round-robin debate with a judge is packaged (debate()); router-selected speakers and open-ended agent chatter deliberately stay build-your-own. The SDK's orchestrator-workers (sub-agents as tools) cover the rest.
Not a fork or patch of laravel/ai. It composes the SDK's public API only, behind a single adapter seam.

License

MIT

timmcleod / agent-workflows

Package info

Security

Statistics