Local Discovery Agent: A Local-First Approach

Updated 6 July 2026

Local Discovery Agents are local-first components that discover and resolve actionable resources such as internal APIs, datasets, and research materials within bounded environments.
They leverage a three-path model—local cache, shared graph, and browser fallback—to ensure low-latency execution, self-correcting verification, and economic efficiency.
Key applications span offline paper retrieval, image geolocation, drug discovery workflows, and secure agent naming, establishing a versatile pattern across domains.

A Local Discovery Agent is a local-first agent component that discovers and resolves actionable resources within a bounded execution context—such as internal APIs, research papers, geospatial evidence, datasets, agent endpoints, or drug candidates—while keeping execution, verification, and policy decisions close to the host, user, cluster, or on-premise environment. In Unbrowse, the term denotes the local agent that resolves intents through a three-path model of local cache, shared graph, and browser fallback, using direct calls to internal or “shadow APIs” rather than browser-first interaction (Tham et al., 1 Apr 2026). Related work uses the same architectural idea for offline paper retrieval, image geolocation, geoscientific archive exploration, parameterized drug-discovery workflows, and secure agent naming and resolution, which suggests that “Local Discovery Agent” has become a general pattern rather than a single domain-specific implementation (Kumar et al., 7 Apr 2026, Li et al., 27 Jan 2026, Pantiukhin et al., 24 Feb 2026, Li et al., 20 May 2025, Rodriguez, 21 Jan 2026).

1. Scope and meanings of the term

Across recent literature, the phrase denotes a locally executed discovery role rather than a single fixed protocol. In Unbrowse, the agent discovers callable first-party interfaces behind websites and executes routes locally with user credentials (Tham et al., 1 Apr 2026). In Paper Circle, the Local Discovery Agent is the offline retrieval worker in the Paper Search Agent, responsible for fast search over locally curated corpora, optional merging with online sources, deduplication, multi-criteria scoring, diversity-aware ranking, and synchronized exports (Kumar et al., 7 Apr 2026). In LocationAgent, the term is effectively instantiated as a hierarchical geolocation agent that performs structured hypothesis–verification cycles with external tools while maintaining coherent local state through the Reasoner–Executor–Recorder architecture (Li et al., 27 Jan 2026). In PANGAEA-GPT, the local discovery role appears in an on-premises Supervisor-Worker system with strict data-type-aware routing and deterministic sandbox execution (Pantiukhin et al., 24 Feb 2026). In AdaptiveResolver and ANS, it appears as a local resolver for agent names, capability records, DIDs, VCs, and policy-checked service selection (Zinky et al., 5 Aug 2025, Mittal et al., 29 Apr 2026).

System	Local discovery role	Core mechanism
Unbrowse	Local intent resolver for internal APIs	Local cache, shared route graph, browser fallback
Paper Circle	Offline retrieval worker	BM25, TF–IDF, deduplication, MMR
LocationAgent	Hierarchical localization agent	RER, clue-exploration tools, strict grounding
PANGAEA-GPT	On-premises discovery-and-analysis worker system	Supervisor-Worker routing, sandboxed code execution
ANS / AdaptiveResolver	Local resolver for agent identity and endpoints	DID/VC verification, policy, context-aware resolution

This diversity clarifies a common misconception: “local” does not mean isolated or purely offline. Several systems are explicitly local-first while still consulting shared graphs, registries, mirrored indexes, or authoritative services when local information is insufficient (Tham et al., 1 Apr 2026, Zinky et al., 5 Aug 2025, Mittal et al., 29 Apr 2026).

2. Unbrowse as a canonical architecture

Unbrowse provides the most explicit design blueprint for a Local Discovery Agent as a reusable architecture. Its starting point is the observation that every modern website already exposes first-party endpoints behind its user interface—XHR/fetch, GraphQL, and WebSocket routes—which the paper calls internal or shadow APIs (Tham et al., 1 Apr 2026). The Local Discovery Agent maintains a local route cache and credential vault, resolves intents via a three-path execution model, passively learns routes during fallback browsing, publishes improvements with delta-attribution, enforces trust and self-correction through verification loops and failure counters, makes economically rational path decisions under x402 fee tiers, and executes routes locally with user credentials while never exfiltrating secrets (Tham et al., 1 Apr 2026).

The Unbrowse route graph stores endpoint definitions, parameter and response schemas, auth descriptors, and derived health, freshness, and confidence signals. Route discovery is instrumented through Kuri, a Zig-native CDP broker, which intercepts XHR/fetch, GraphQL, and WebSocket traffic; decompresses gzip and deflate; normalizes charset; strips tracking query parameters; canonicalizes URL paths by replacing dynamic segments with typed placeholders; infers query and body schemas; and records auth descriptors such as bearer headers, cookies, CSRF tokens, and refresh workflows (Tham et al., 1 Apr 2026). The local data model is explicit: a RouteNode includes endpoint, method, query/body/path parameters, headers, auth_descriptor, response_schema, freshness, ttl, reliability, verification_status, consecutive_failures, last_verified, confidence, version, and attribution (Tham et al., 1 Apr 2026).

A distinctive feature is the packaging of learned routes as reusable skills. When a route is learned during fallback browsing, the agent can publish a package that includes SKILL.md, api.ts, and local-only auth.json, with credentials retained only in the local encrypted vault (Tham et al., 1 Apr 2026). This makes the Local Discovery Agent both an execution engine and a contributor to a collectively maintained discovery graph.

3. Execution logic, self-correction, and economic discipline

Unbrowse formalizes local discovery through a three-path execution model: local cache, shared graph, and browser fallback (Tham et al., 1 Apr 2026). Path 1 executes a matching, fresh, verified route from the local cache, with a default TTL of 24 hours. Path 2 queries the shared graph through semantic search, potentially paying Tier 3 for lookup and Tier 1 for local installation of the skill. Path 3 launches headless Chrome via Kuri, captures HAR or network events, extracts new routes, verifies them, installs them locally, and publishes the learned skill; for sites without detectable APIs, structured data may be extracted from HTML (Tham et al., 1 Apr 2026).

The system is explicitly self-correcting. Freshness is defined as

$\text{freshness} = \frac{1}{1 + d/30},$

where $d$ is days since last verification, and routes stale beyond 24 hours are prioritized for re-verification (Tham et al., 1 Apr 2026). Every 6 hours, the verification loop replays safe GET endpoints, compares live responses with documented schemas, updates health and reliability, and flags drift. On endpoint failure or schema mismatch, the agent increments consecutive_failures; if trust falls below threshold, it marks the route deprecated or disabled, triggers browser fallback, relearns the route, replaces the local version, and publishes the delta (Tham et al., 1 Apr 2026). The composite graph-search score uses 40% embedding similarity, 30% reliability, 15% freshness, and 15% verification status (Tham et al., 1 Apr 2026).

Unbrowse also ties adoption to explicit economic reasoning. It uses a three-tier x402 micropayment model: Tier 1 is a one-time install fee for discovery documentation, Tier 2 is an opt-in per-execution fee for site owners, and Tier 3 is a per-query search fee, typically $0.001–0.005 per query (Tham et al., 1 Apr 2026). The necessary adoption condition is that the agent uses the shared graph only when the total fee is lower than the expected browser rediscovery cost, which is decomposed into latency, compute, token usage, and retry cost under failure (Tham et al., 1 Apr 2026). This economic outside option is central to the system’s claim of voluntary and self-correcting participation.

4. Retrieval, ranking, and indexing variants

Outside Unbrowse, Local Discovery Agents frequently appear as specialized retrieval-and-ranking systems. GRAIL is a real-time agent discovery framework that combines an inverted tag index, a dense context index built through pseudo-document expansion, and a fine-grained example matrix for MaxSim resonance (Xu, 4 May 2026). Its online path predicts capability tags with a LoRA-tuned Qwen2.5-3B-Instruct SLM, runs sparse and dense recall in parallel, merges candidates, and reranks them with

$S_{\text{final}}(q, a) = \alpha [f(q) \cdot v_{\text{ctx}}^{(a)}] + (1-\alpha) r(a,q),$

where $r(a,q)$ is the maximum similarity between the query and the agent’s usage examples (Xu, 4 May 2026). On AgentTaxo-9K, GRAIL reports 394.48 ms average end-to-end latency, compared to 31,186 ms for LLM-Router, and achieves $R@10 = 91.45\%$ (Xu, 4 May 2026).

Paper Circle instantiates the pattern for academic literature discovery. Its Local Discovery Agent is the offline retrieval worker in a coder-LLM multi-agent framework and operates over locally curated JSON corpora with BM25 and TF–IDF indexes, optional sentence-transformer embeddings, optional Qwen3-Reranker-0.6B reranking, and offline or mixed offline/online retrieval (Kumar et al., 7 Apr 2026). It computes similarity, recency, novelty, normalized BM25, and optional citation features, then combines them through mode-specific weights and applies Maximal Marginal Relevance:

$\mathrm{MMR} = \arg\max_{p \in R \setminus S} \left[ \lambda \cdot \mathrm{sim}(p,q) - (1-\lambda)\cdot \max_{s\in S}\mathrm{sim}(p,s) \right]$

(Kumar et al., 7 Apr 2026). The system persists synchronized papers.json, links.json, stats.json, summary.json, and retrieval_metrics.json, and exports CSV, BibTeX, Markdown, and HTML after every step (Kumar et al., 7 Apr 2026).

These systems share a structural pattern: local indexes, explicit ranking formulas, and deterministic or semi-deterministic artifact generation. A plausible implication is that “local discovery” increasingly refers not just to locality of deployment, but to locality of ranking state, reproducible outputs, and bounded retrieval loops.

5. Hypothesis–verification loops and structured memory

A second major lineage treats the Local Discovery Agent as a disciplined loop that separates reasoning from factual verification. LocationAgent formalizes image geolocation as abductive constraint satisfaction over state $S_t = \{L_t, E_t\}$ , where candidate regions and verified evidence evolve through a reasoning–verification–projection cycle (Li et al., 27 Jan 2026). Its RER architecture—Reasoner, Executor, Recorder—keeps hierarchical reasoning inside the model while delegating environmental, infrastructure, semantic-symbol, and image-matching evidence to tools (Li et al., 27 Jan 2026). On CCL-Bench, it reports 52.33% accuracy at 1 km, 82.00% at 25 km, and 84.67% city-name accuracy, with 100.00% Location Compliance (Li et al., 27 Jan 2026).

DrugPilot applies the same principle to drug discovery. Its Local Discovery Agent is built around a Parameterized Memory Pool, where concise keys map to structured values such as SMILES lists, sequences, matrices, and files, enabling tool calls without forcing large parameters into the LLM context (Li et al., 20 May 2025). The system supports eight drug-discovery tasks and uses the Feedback and Focus mechanism to correct function selection and parameter formatting errors (Li et al., 20 May 2025). On its drug-discovery tool instruction dataset, it reports task completion rates of 98.0% on simple tasks, 93.5% on multiple-function tasks, and 64.0% on multi-turn tasks (Li et al., 20 May 2025).

PANGAEA-GPT generalizes the pattern to heterogeneous Earth-science archives through a centralized Supervisor-Worker topology, strict data-type-aware routing, deterministic sandboxed execution, and execution-feedback-based self-correction (Pantiukhin et al., 24 Feb 2026). CORAL extends it to open-ended discovery by using persistent attempts, notes, and skills, plus heartbeat-based reflect, consolidate, and pivot interventions (Qu et al., 2 Apr 2026). InternAgent-1.5 pushes further toward continuous scientific discovery with Generation, Verification, and Evolution subsystems supported by Strategy-Procedural Memory, Task-Episodic Memory, and Semantic-Knowledge Memory (Feng et al., 9 Feb 2026).

These systems clarify another misconception: a Local Discovery Agent is not merely a search interface. In many domains, it is a long-horizon controller that couples local memory, tool execution, and explicit verification.

6. Naming, trust, and locality as security primitives

A third lineage focuses on local discovery as secure name resolution and capability verification. The agent:// URI scheme decouples identity from topology through a trust root, a hierarchical capability path, and a sortable unique identifier, enabling capability-based discovery through trust-root-scoped DHT key derivation:

$k = \mathrm{SHA256}\!\left(\text{canonical}(T)\, \|\, "/" \, \|\, \text{canonical}(C)\right)$

(Rodriguez, 21 Jan 2026). Discovery can operate in exact or prefix modes, and cryptographic attestation uses PASETO v4.public tokens whose claims bind issuer, subject URI, validity interval, and authorized capability prefixes (Rodriguez, 21 Jan 2026). The paper reports 100% capability-path coverage on 369 production tools, $F1 = 1.0$ on a 10,000-agent discovery benchmark, and microsecond-scale parsing, canonicalization, key derivation, and verification operations (Rodriguez, 21 Jan 2026).

ANS translates similar ideas into Kubernetes. Its Local Discovery Agent is a cluster-local caching resolver that consults an ANS Registry, resolves DIDs and Verifiable Credentials, applies OPA discovery policies, and returns vetted endpoints (Mittal et al., 29 Apr 2026). In the proof-of-concept environment, reported means are 45 ms for agent registration, 12 ms for agent discovery, 78 ms for capability verification, 3 ms for policy evaluation, and 15 ms for certificate validation; sub-10 ms responses were observed in demonstrated service paths, and a 50-agent workflow simulation completed successfully in the scripted setup (Mittal et al., 29 Apr 2026).

AdaptiveResolver defines a closely related Local Discovery Agent as a microservice that ingests Agent Fact cards, listens to local discovery signals such as mDNS, SSDP, peer beacons, gossip, and edge directories, scores candidate endpoints with a context-aware utility function,

$U(E|C) = \sum_{i=1}^{n} w_i \,\tilde{x}_i(E,C),$

and optionally negotiates trust and QoS before returning a Tailored Response (Zinky et al., 5 Aug 2025). Earlier infrastructure work already used a host-resident discovery/control agent in LISA, which monitored host, network, and application state, streamed data to MonALISA, and enacted remote tuning of TCP, NIC, and FDT parameters during large-scale WAN data transfers (Dobre et al., 2011).

Taken together, these papers show that locality can be a security primitive: it bounds trust decisions to local caches, local policy engines, local certificates, or local discovery signals before wider federation is attempted.

7. Performance characteristics, limits, and open issues

Published results show that Local Discovery Agents can materially reduce latency, improve retrieval quality, and increase structured task success, but they also expose distinct failure modes. In Unbrowse’s single-host live-web benchmark across 94 domains, fully warmed cached execution averaged 950 ms versus 3,404 ms for Playwright, with 3.6× mean speedup and 5.4× median speedup; on 18 well-cached domains, routes completed in under 100 ms, with the fastest case at 79 ms versus 2,289 ms (Tham et al., 1 Apr 2026). Cold-start median was 8,200 ms, mean 12,400 ms, 90th percentile 22,000 ms, and skill publish success was 18/20, with cached-call amortization typically breaking even after 3–5 uses (Tham et al., 1 Apr 2026). GRAIL reaches sub-400 ms agent retrieval, but relies on a 3B SLM that may not meet the same target on CPU-only devices (Xu, 4 May 2026). Paper Circle’s reranking can improve MRR, but BM25 plus reranker is reported as approximately 28× slower than lighter configurations (Kumar et al., 7 Apr 2026).

The limits are equally consistent across domains. Unbrowse notes anti-automation detection lowers speedups on protected sites and only plans robots.txt- and WAF-aware policies (Tham et al., 1 Apr 2026). GRAIL identifies taxonomy drift and unreported P50/P95 latency distributions as open issues (Xu, 4 May 2026). Paper Circle highlights domain drift, stale local corpora, metadata inconsistency, and API failures (Kumar et al., 7 Apr 2026). LocationAgent remains sensitive to ambiguous scenes, OCR noise, and incomplete retrieval coverage (Li et al., 27 Jan 2026). HLCD shows that local equivalence classes can still interfere with orientation unless explicit score comparisons are used to separate V-structures from equivalent local patterns (Ling et al., 2024).

A plausible implication is that the research frontier is moving toward hybrid systems that combine local execution, shared but inspectable indexes, explicit verification loops, and stronger trust layers rather than toward purely monolithic, browser-first, or end-to-end generative designs.