# Agent Forge: product and business thesis

Research date: October 21, 2026 (America/New_York). This is a decision document, not a forecast, launch announcement, or claim of product-market fit. Vendor pages are current observations and can change. No customer interviews, paid pilots, revenue, or market-size study have been completed.

## The decision

Build a visual operations workspace for people who already run agent teams. Start with small automation agencies managing several client workflows. Their proposed job-to-be-done: understand what is happening, find the next decision, and deliver a client-readable account of the work without exposing private prompts or asking the client to read developer traces. This audience and pain are hypotheses to validate.

Our present advantage is a working combination of original spatial UI, persistent state, native orchestration and an external LangGraph adapter. It is not yet a defensible moat. A strong product must measurably improve comprehension and intervention while preserving the user's runtime and provider choices. We cannot honestly estimate a probability of commercial success today.

## What the research says

| Alternative | Observed capability / offer | Implication for Forge |
|---|---|---|
| [Pixel Agents](https://github.com/pixel-agents-hq/pixel-agents) | Existing office characters, activity, hooks, permissions, editable layouts, local extension/CLI; broader providers on its roadmap | Characters and portability are not unique claims. Test agency operations as a narrower job, rather than trying to win a furniture contest. |
| [Star Office UI](https://github.com/ringhyacinth/Star-Office-UI) | Multi-agent pixel status world and customization; code and art have different reuse terms | Free visual substitutes exist. Keep original assets; charge for repeated operational value. |
| [Agents Office](https://sahni.ai/agentsofficev3/) | 3D departments, task routing, local model login and note-linked results in a free beta | A pretty office plus task execution is already offered. |
| [LangSmith](https://www.langchain.com/pricing) | Free developer tier; Plus $39/seat/month plus usage; tracing, evaluation and deployment platform | Mature engineering tools already address debugging. Connect or deep-link to them instead of rebuilding their entire stack. |
| [Langfuse](https://langfuse.com/pricing) | Free tier, Core $29/month, Pro $199/month; usage and team controls vary | Low-cost observability creates price pressure. Event/trace units and retention differ; sticker prices are not equivalent workloads. |
| [AgentOps](https://www.agentops.ai/) | Cross-framework traces, costs and replay; Pro starts at $40/month | Cross-framework monitoring by itself is also occupied territory. |
| [CrewAI](https://crewai.com/pricing) | Free basic workflow offer and custom enterprise governance/deployment | Avoid assuming enterprise buyers need another general orchestration platform. |

These pages show available substitutes and pricing offers, not how much users actually pay or why they renew. Repository activity is evidence of interest, not purchasing intent. The inference is that the interface is attractive enough to explore, but a visual wrapper alone faces substantial free competition.

## A meaningful wedge to test

**Client-safe mission rooms for agencies running mixed agent teams.** The room is an accessible view of work, not a simulation of productivity. Its useful objects should represent concrete things:

- A desk shows an identified runtime node, current assignment, last confirmed activity and supported controls.
- A handoff line shows a real dependency, including parallel branches and joins.
- A decision inbox shows explicit approval requests, failed work or lost reporting, not a made-up confidence score.
- A deliverable contains the output, its source links, generating run and available usage measurements. It does not pretend to expose hidden chain-of-thought.
- A client portal eventually shows only the fields the agency chose to publish. Private inputs and raw prompts stay excluded by default.
- A focused list view and keyboard controls carry the same information as the world. Motion, sound and cosmetics are optional.

None of this is assumed exclusive. The differentiation would be a reliably simpler agency-to-client workflow and superior onboarding. Verify it against both a plain status table and an existing office tool. If people enjoy watching but do not return for operational work, narrow the product to a local creative tool rather than continuing a SaaS story.

## Built now, then next

**Built locally:** persistent JSON team import, original top-down world, themes and outfits, node inspection, event ingestion with scoped hashed tokens and ordering checks, result display, an anonymous evidence-only brief with missing/conflict handling, native RQ execution, an opt-in synchronous LangGraph node adapter, and a paired outbound runtime connector with world-launched task handlers, persisted run history, cooperative controls and acknowledgements. The four-node external mission performs evidence review, parallel pricing/product analysis, then an Overseer review using real bounded model calls. LangGraph owns execution; Forge observes it. The adapter exports selected fields only and does not scrape credentials or upload executable agents.

**Not built as a production service:** customer identity, enforced membership-based tenant access, client publishing, billing, signed expiring viewer links, durable external command acknowledgement, universal cross-framework pause/resume, scalable event storage, migrations, recovery guarantees, enterprise certifications or an SLA. A shared API key is not tenant authorization. A heartbeat proves a process is reporting; it does not prove useful progress or correct results.

| Stage | Work | Gate before expanding |
|---|---|---|
| Days 1–14: establish usefulness | Package LangGraph adapter; install guide; dependable failures/results; attention board; compare world against table; interview 10 operators about the last real incident and last client handoff | 3 external teams connect their own workflow; median setup under 15 minutes; 4/5 observed users correctly identify working, failed and disconnected states without coaching |
| Days 15–30: prove paid pain | Concierge pilots; read-only client report prototype; make export/redaction explicit; add the next adapter only when pilot users request it (CrewAI is a candidate) | At least 2 of 3 design partners pay an actual agreed fee and use it weekly for four weeks; no fabricated commitments or automatic outreach |
| Days 31–60: earn hosted beta | Identity, tenant isolation tests, field-level read models, expiring tokens, retention, append-only event model, quota/rate limits, backups and restore drill; acknowledged commands only for adapters that implement them | Isolation/security review passes; no private prompt exposure in client fixtures; bounded event volume and support load; measurable reduction in time-to-find a blocker |
| Days 61–90: repeatability | Self-serve install, first paid workspace plan, usage metering, agency report delivery; compare activation and renewal by cohort | Users renew because it saves work; onboarding succeeds without founder help; gross margin and support cost justify acquisition spend |

Proposed usability threshold: on equivalent tasks, reduce median time to identify a blocker by at least 30% versus a simple table, with no loss of accuracy. This is a target, not a measured result. Use a counterbalanced test with at least five participants first, then repeat with more users; record errors as well as speed. Prefer concrete interviews and paid pilots to a broad market-size deck.

## Monetization experiments

Start with a paid integration/design-partner engagement: test $300–$1,000 one-time for a tightly scoped installation, workflow mapping and feedback sessions. Record the labor; service revenue is not software retention. Do not promise a production deployment before access controls are ready.

Then test recurring prices, not public commitments:

| Proposed plan | Price hypothesis | Reason to pay, if built and validated |
|---|---:|---|
| Local / community | Free | One local world, essential connection and inspection; a low-friction acquisition path |
| Builder | $29/month | Convenient history, useful alerts and saved worlds |
| Studio | $79/month | Several worlds, collaboration and longer history |
| Agency | $199/month | Client-safe branded rooms, permissioned approvals and handoff reports |
| Private deployment | Scoped quote after requirements | Deployment, operation and support costs; no invented security certifications |

Prefer workspace/client-world allowances with transparent event and retention limits. Heartbeats should not create surprising bills. Start with hard caps and optional upgrades rather than automatic overages. Keep inference on the customer's provider account initially. Never advertise unlimited agents, unlimited events, guaranteed savings or included inference before measuring cost.

Cosmetic packs, branded environments and licensed designer assets can be secondary revenue. They should not fund the core thesis until users demonstrate willingness to pay. A marketplace adds licensing, quality, creator payouts, support and moderation work; postpone it. Avoid percentage-of-savings pricing until attribution is credible. Enterprise procurement is a later motion, not a shortcut to high prices.

## Economics: assumptions we can change

Illustrative monthly model, not a forecast: 50 Builder accounts at $29, 15 Studio at $79, and 5 Agency at $199 yields **$3,630 MRR** before churn, tax, refunds or collection costs. Assume shared infrastructure $150, metered storage/egress $140, support 12 hours at an economic cost of $40/hour, and payment cost 3% + $0.30 per 70 invoices ($129.90). Contribution is **$2,730.10 (75.2%)** before product development, acquisition, general overhead and tax. These cost assumptions are not vendor quotes or verified operating costs. Reaching that customer count is unproven.

At $29/month, an account requiring one support hour per month at $40/hour is already unattractive before infrastructure. Measure onboarding support per account, active-world density, event bytes, retention cost and weekly meaningful use. Estimate CAC only after a real acquisition experiment. Do not calculate confident LTV from a few weeks of churn data. A public founder audience is useful distribution, but music-site visits are not evidence of agent-tool buying intent.

## Hosting with KU13SH

The current KU13SH source and deployment script describe a static S3-to-Amplify release. The live site resolves through CloudFront. This can host a marketing page or static read-only showcase; it does not automatically run FastAPI, Postgres, Redis or long-lived workers. [AWS distinguishes frontend hosting and backend resources](https://docs.aws.amazon.com/amplify/latest/userguide/deploy-backend.html).

Recommended sequence: keep this working live integration local; prepare a marketing/showcase path at `www.ku13sh.com/agent-forge/`; put an authenticated application on a separate origin such as `forge.ku13sh.com` once ready. A separate backend service owns the API, event store and queue. External runtimes connect outbound over HTTPS; the server never receives local ChatGPT/Codex credentials. Allowlist the exact application origin. Keep the personal site deployment separate so other projects and unpublished assets cannot accidentally ship with Forge.

A future paid hosting choice needs an approved spending cap and measured load; no infrastructure was purchased or provisioned in this pass. No public deployment, DNS change, existing-site edit or shared exposure of the local Codex account is needed to watch this run. Before publishing a client-facing beta: membership authorization, token rotation/revocation, origin checks, telemetry quotas, backups, deletion/retention controls, secret filtering, and an independent isolation review are release requirements tied to the actual architecture.

## Techniques worth adopting

1. **Capability negotiation:** each adapter declares observe/results/usage/approve/pause/resume/cancel separately. No working button without an acknowledged implementation. A graph node boundary is not an immediate process stop.
2. **Interoperability:** evaluate [AG-UI](https://docs.ag-ui.com/introduction) for runtime-to-frontend events and [OpenTelemetry GenAI conventions](https://opentelemetry.io/docs/specs/semconv/gen-ai/) for trace identity. Pin schema versions, test mappings and keep our simple internal contract; do not label today's adapter compliant before conformance tests.
3. **Human checkpoints:** use [LangGraph interrupts](https://docs.langchain.com/oss/python/langgraph/interrupts) with durable checkpoints and idempotent side effects when adding approvals. Never replay paid or external actions just because a heartbeat is old.
4. **Framework events:** [LangGraph streaming](https://docs.langchain.com/oss/python/langgraph/streaming) and [CrewAI listeners](https://docs.crewai.com/en/concepts/event-listener) are explicit sources of lifecycle events. Our first sync wrapper is intentionally smaller; it does not claim universal nested-agent/tool tracing.
5. **Outcome receipts:** distinguish generated, reviewed, approved and executed artifacts. Record usage as reported; leave dollar costs unknown when no price exists. A completed LLM call is not validated business work.
6. **Attention rather than surveillance:** make stale feeds and blocked decisions visible. Avoid motivational theater or claims that pressuring an agent makes it productive. The Overseer's job is bounded routing, review and escalation.

## What would change the thesis

Stop or reposition if users prefer free office tools, cannot connect within 30 minutes, refuse a paid pilot, or only open the room for novelty. If nontechnical clients understand the work better but developers prefer LangSmith, that supports a client presentation layer and integrations. If only solo builders pay, pursue a low-support local product. If governance dominates interviews, address identity and controls before more environments. The best next investment follows observed use and renewals, not the largest possible feature list.


## User-feedback follow-up and current control scope

See [USER_FEEDBACK.md](USER_FEEDBACK.md) for the review sample, positive reactions, reported failures and the maintainer’s reported fix. The local prototype now implements pairing, registered task launch, run history and cooperative pause/resume/cancel through the outbound Python connector. Commands distinguish requested and acknowledged states. A lost lease is interrupted without automatic replay. These controls require wrapped step boundaries and a live process; they do not guarantee immediate termination or recovery after restart. Public client permissions and hosting remain separate work.