• Member
• November 14, 2026

Crowdsourced Open Source Recipient Side Agents

Proposal: Crowdsourced Open-Source Recipient-Side Agents

Open Recipient Agent Commons (Swa-Pratinidhi) · Srijan Sanchar · Shailendra Jaiswal · 14 November 2026

Summary

Software now decides what advertising and content reach people, and almost all of that software works for the sender. Ad platforms run AI agents that sell attention, OS-level assistants are becoming the gate to every screen, and content and advertising are merging into fragments and serial ads. Laws such as India’s DPDP Act are tightening, but compliance on paper is not protection in practice: a 2026 audit found TikTok formally compliant with EU rules while still profiling minors through promotional content.

Srijan Sanchar proposes an Open Recipient Agent Commons, working name Swa-Pratinidhi, “one’s own representative”. It is a free, open-source AI agent that runs on a person’s own device and acts only for them. It filters ads and content by the person’s declared needs, explains why anything reached them, negotiates offers and issues a receipt, protects children and blocks manipulation, and, with consent, contributes anonymised evidence to independent audits. It builds on Srijan Sanchar’s Open algorithms for attention framework.

Everything that decides how the agent acts is open: code, schemas, reference rankers, on-device models, test sets and assurance criteria. Builds are signed and verifiable so fake agents can be told apart. The crowd (citizens, parents, developers, researchers, language volunteers, creators and red-teamers) improves it, rewarded through a contribution ledger where benefit follows contribution and control does not follow money. A consortium of universities stewards it, with a stewardship council on which platforms hold at most an observer seat, and no advertising money funds the agent.

The promise is an agent every person can check and that costs them nothing, with fairer reach for creators and small sellers, and a public-interest counterweight to closed platform assistants, built like UPI. The main challenges are distribution, platform lockout, fake agents, attacks hidden in ads, and funding without capture. The first steps are to publish the proposal and schema, sign up three to five founding universities in early 2027, and pilot a minimal agent with Katha Setu readers and a TAI children’s-content panel before India’s children’s-data obligations apply in May 2027.

1. Context

Advertising and content are now decided by software on the sender’s side, and the receiver has no equivalent.

•           Agents trade with agents. Ad platforms run AI sellers and buyers; one Indian seller alone reports 25 trillion predictions a day.

•           Content and ads are merging. Fragments, serial ads and AI-inserted placements blur what counts as an advertisement.

•           Orchestrators control access. Feeds, and increasingly OS-level assistants, decide what reaches a person and capture most of the value.

•           Attention is the scarce resource. Value of a moment ≈ attention × capability to purchase × trust.

•           Paper compliance is not protection. A 2026 audit of TikTok found formal compliance with the EU ban on profiled ads to minors, yet promotional content profiled to minors 5–8 times more strongly than formal ads to adults (KInIT, FAccT 2026).

•           India has the rails. UPI handled 25.47 billion transactions in October 2026. The DPDP Rules bring consent managers from November 2026 and children’s-data obligations from May 2027.

Srijan Sanchar’s foresight places the ad-funded free model in a pre-threshold state, five to ten years from breaking (Structural Threshold Foresight). The question for that transition is whose agent orchestrates what reaches the receiver. Closed assistants built by platforms will answer it by default unless an open alternative exists.

2. Imperatives

For a recipient agent to protect people rather than become another gatekeeper, eight things must be true.

1.         Loyal to the person, provably. Its objective is declared and its code, rankers and models are open to inspection.

2.         Local-first. Data and decisions stay on the person’s device; only what the person chooses is shared.

3.         Open the rules before personalising the agent. It runs on open standards and public rails, never inside one platform’s walls.

4.         Explainable in every Indian language. It can say why anything was shown, accepted or refused.

5.         Hard to fool. It is protected against prompt injection and “agent SEO” hidden in ads and product pages.

6.         Safe for children by design. A parent-controlled mode, no profiling, and protection in content as well as ads.

7.         Independently assured. It is certified by people who answer to receivers, not by those who build or fund it.

8.         Funded without selling the person. It never earns money from the data it protects.

The Dharma gate applies to every release: would the person, fully informed, endorse how their agent acted for them?

3. Purpose

The purpose is to give every person in India, and later anywhere, a free, open, verifiable agent that acts only for them when advertising, offers and content try to reach them. The crowd improves it and a neutral university consortium stewards it.

Open Recipient Agent Commons: crowd, commons, consortium, agent

The crowd contributes rankers, labels, tests and translations to an open commons. The consortium reviews, certifies and signs releases. The agent runs on the person’s device, receives offers and content from platforms, and returns only consented, anonymised traces that improve the commons.

The agent has five jobs.

1.         Filter: apply the person’s own ranking (by declared intent, not platform engagement) to ads, offers and feeds.

2.         Explain: show why each item reached the person, in their language, using signed reason codes where platforms provide them.

3.         Negotiate: post time-boxed intent windows, collect offers and accept the best, issuing a best-execution receipt.

4.         Protect: refuse manipulation, block profiling of children, flag unlabelled ads and scams.

5.         Witness: keep a private attention mirror for the person and, with consent, contribute anonymised evidence to independent audits.

It builds directly on Srijan Sanchar’s Open algorithms for attention (the Open Attention Stack, Attention Trace Schema and reference rankers), and is the person-side half of the Software-Based Advertising foresight.

4. Open Source

Everything that decides how the agent acts for a person is open. That makes its loyalty verifiable, and it reaches the top rung of the Open Algorithms ladder, which platform rankers cannot.

What is opened	Why	Suggested licence
Agent code (filtering, explanation, negotiation, protection)	Anyone can inspect it for hidden tolls or data leaks	Apache 2.0
Schemas: Attention Trace Schema plus an advertising and offer extension (ATS-Ad), reason-code codebook, best-execution receipt	Shared language between agents, platforms and auditors	CC BY 4.0
Reference rankers (R0 chronological, R1 declared intent, R2 commonality-aware, R3 track-record-weighted)	People can choose or compare how they are ranked	Apache 2.0
On-device models (small, open-weight, Indian-language capable)	A closed model underneath would leak loyalty	Open-weight licence with a use policy
Test and red-team sets: manipulation patterns, unlabelled ads, scams, prompt injections	Anyone can check the agent works	CC BY 4.0, with privacy screening
Assurance criteria (Dharma Assurance)	Certification can be inspected and challenged	CC BY 4.0

Open code must mean an honest install. Releases are built reproducibly, signed by the consortium and checkable on the device, so fake “protective” agents can be told apart. Anyone may fork the code, but only certified builds may use the protected name and mark.

What stays closed: the person’s data. Raw traces never leave the device. Only consented, aggregated, privacy-protected evidence reaches the commons.

Security by openness. Open test sets and public red-teaming find weaknesses faster than one company could, especially prompt-injection attacks hidden in ads.

5. Crowd Sourcing

The crowd improves what no single team can: knowledge of local manipulation tactics, languages, scams and preferences across India. Benefit follows contribution; control does not follow money.

Contributor	What they contribute	Example
Citizens and parents	Report manipulative or unlabelled ads; donate consented traces; rate explanations	“This ‘free trial’ ad hid a subscription”
Developers	Code, rankers, integrations with UPI and ONDC apps, security fixes	A ranker tuned for farmers’ purchases
Researchers and students	Audits, indices, evaluation studies, model fine-tuning	A study of ad load on low-income users
Language volunteers	Reason-code translations, dialect testing	Explanations in Bhojpuri or Tulu
Creators and small sellers	Report unfair ranking; test offer formats; supply verified product data	A toy maker flags hidden self-preferencing
Red-teamers	Attack the agent with prompt injection, fake offers, dark patterns	An ad that tries to instruct the agent to accept it

Incentives without corruption. A contribution ledger records standing for every accepted contribution. Value realised later (grants, licensing, assurance fees) is shared by standing, with no per-view payment, standing that cannot be sold, and no votes for money. Students earn academic credit, researchers get publishable data, and small sellers get fairer reach.

Quality and integrity.

•           Every contribution is reviewed by maintainers and, for sensitive areas (children, health, finance), by domain experts.

•           Integrity checks catch duplicates, bots and coordinated manipulation.

•           Reputation is earned, and reviewers’ own contributions are checked by others.

•           Nobody can contribute ads, sponsorship or paid ranking into the commons.

6. Consortium of Universities

Universities are the right stewards because they are neutral, trusted, multidisciplinary and renewed every year by students. A consortium also spreads the work across India’s languages and regions. India already has university-led open AI efforts in Indian languages, such as AI4Bharat at IIT Madras and the BharatGen consortium led by IIT Bombay.

Role	Who (illustrative)	Responsibility
Convening anchor	One lead university, with Srijan Sanchar as originator and secretariat	Charter, coordination, releases
Technical labs	Computer science and AI departments	Agent, rankers, on-device models, security
Law and policy	National law universities	DPDP compliance, consent design, right to delegate, assurance rules
Behaviour and society	Psychology, sociology, economics, media studies	Manipulation patterns, fairness audits, child safety
Language	Regional universities and language institutes	Explanations and testing in Indian languages
Regional and teaching partners	Universities such as IIMT University, Meerut	Student projects, local pilots, curriculum
International partners	Universities in open-source and AI-ethics research	Cross-checks, global standards, comparison with other regimes

Governance.

•           A stewardship council with defined seats: citizens and parents, creators and small sellers, researchers, member universities and child-safety experts. Platforms and advertisers hold at most an observer seat.

•           Separation of roles: those who build a release cannot certify it; an independent assurance panel does.

•           Open minutes, an open roadmap and a public register of conflicts of interest.

•           Ledger standing gives a benefit share, not votes.

Funding. Public research grants and national AI programmes, CSR and philanthropic funds, and fees for independent assurance of third-party agents. No advertising or platform money funds the agent itself; any platform contribution goes only to a blind pooled fund for audits.

What universities gain. A living laboratory for AI, law and society; real data under ethical governance; student projects with public impact; and a share in a national public good.

7. Promise

To every person: an agent that works for you, that you can check, and that costs you nothing.

For	The promise
People	Fewer manipulative ads, clear reasons for what you see, better offers on your terms, data that stays with you
Parents and children	A child mode that blocks profiling and persuasion in ads and content
Elderly and first-time users	A scam shield that explains risk in their own language
Creators and small sellers	Fair, explainable reach: found because you fit someone’s need, not because you paid the gatekeeper
Advertisers who play fair	Trusted, consented intent, which converts better than surveillance
Universities	A national living lab and a role in shaping AI for the public good
India	A public-interest counterweight to closed platform assistants, built like UPI: an open rail with many apps on top

In Srijan Sanchar’s frameworks, the commons is a structural intervention (Structural Threshold Foresight): it changes who orchestrates rather than propping up the walled model. It is the only path that keeps Integrity, Impact and Vitality in balance (HEM), and it moves the CRRAAAS Agent gate to the person’s side.

8. Possibilities

Possibility	What the agent does	Pilot partner (illustrative)
Shopping offers	Posts an intent window (“school bag under ₹800, this week”), collects offers, accepts the best with a receipt	An ONDC buyer app
Feeds and content	Ranks by declared intent, shows an attention mirror, flags serial-ad hooks	Katha Setu readers
Child mode	Blocks profiling, labels every ad, parent approves purchases	Children’s-content panel with TAI
Toys and gifts	Checks BIS certification; lets relatives gift a toy or story without using the child’s data	TAI members
Peer-to-peer trade	Verifies vehicle registration (Vahan) or phone IMEI, holds UPI escrow, checks seller reputation	ONDC peer-to-peer pilot
Scam shield	Spots fake offers and impersonation, explains risk in local languages	Elders’ groups, banks
Farmers	Compares input and equipment offers, flags inflated prices	Farmer producer organisations
Creators	Shows creators how agents rank them and why; supports fragment royalties	Creator collectives
Public audits	With consent, pools evidence for independent audits of platforms	Consortium research panel

Longer-range possibilities: a national right to delegate to an agent of one’s choice; ONDC and UPI apps shipping a certified recipient agent by default; and India exporting the model as a digital public good, as it has with UPI.

9. Challenges

Challenge	Why it matters	Mitigation
Distribution	Most people use the assistant their phone ships with	Partner with UPI apps, ONDC buyer apps and phone makers; seek a public-interest default option
Platform lockout	Platforms can block agents through terms or APIs	Advocate a legal right to delegate and interoperability rules
Fake agents	Malicious copies can pose as protective	Reproducible signed builds, a protected mark, on-device verification
Attacks on the agent	Ads may hide prompt injections and “agent SEO”	Open red-teaming, injection hardening, separation of content from instructions
Capability gap	Platform AI is far larger	Shared standards, pooled crowd intelligence, narrow well-tested tasks
On-device limits	Low-cost phones have limited compute	Small open-weight models; optional trusted processing without raw data
Funding and capture	Money can bend priorities	No ad or platform money for the agent; seats not for sale; open accounts
Privacy of traces	Viewing data reveals religion, politics, health	Local-first, tiered access, aggregation thresholds, legal review before pilots
Children	Highest risk, strictest law (DPDP from May 2027)	Parent-controlled mode, no profiling, independent child-safety review
Consortium coordination	Universities move slowly and differently	Small executive team, clear charter, Srijan Sanchar as secretariat, rotating chair
Adoption and usability	Protection is invisible until it fails	Visible wins first: scam shield, child mode, best price found

First steps (Thought–Action Network: exploratory first, avoid irreversible bets).

1.         Committed now: publish this proposal and the ATS-Ad schema extension on srijansanchar.com.

2.         Provisional, early 2027: sign a memorandum with three to five founding universities; form the stewardship council.

3.         Exploratory, 2027: build a minimal agent (explain, filter, child mode) and pilot it with Katha Setu readers and a TAI children’s-content panel before May 2027.

4.         Provisional, 2028: add negotiation and best-execution receipts through one ONDC buyer app, and a peer-to-peer pilot on vehicles.

Hypotheses to test: people understand signed, localised explanations; the child mode measurably reduces profiled exposure; crowd red-teaming finds injection attacks faster than an internal team; small sellers gain fairer reach under intent-based ranking.

Sources

•           Srijan Sanchar, Open algorithms for attention, October 2026 — https://www.srijansanchar.com/Blogs/Detail/STORIES/Open-algorithms-for-attention

•           Srijan Sanchar, Software-Based Advertising foresight compendium and Assurance note, November 2026

•           KInIT, The DSA’s Blind Spot: algorithmic audit of advertising and minor profiling on TikTok, ACM FAccT 2026

•           NPCI UPI data for October 2026, via Free Press Journal

•           S.S. Rana & Co., Digital Personal Data Protection Rules, 2025

•           Srijan Sanchar frameworks: CRRAAAS, Structural Threshold Foresight, HEM, Thought–Action Network (srijansanchar.com/blogs)

•           Online version: https://claude.ai/artifact/PEahxh6w9BrnBQM4yu4fra

University names are illustrative, not commitments. Mentions of AI4Bharat and BharatGen are from general knowledge and should be checked before external use. This proposal is not legal advice.