Agentic retrieval techniques that hold up in production
Six agentic retrieval techniques backed by 2026 benchmarks: read enforcement, query decomposition, routing, split routers, verification, and step budgets.
Every agent you work with,
reading and writing to the same place.
Escape the lock-in. Wire works with Claude, Codex, and every agent with MCP.
Switch between agents without missing a beat.
One container, read and written from wherever you or your team happen to be working.
claude — ~/competitor-intel
❯ did northwind change their pricing?
⏺ WebFetch(northwind.com/pricing)
⎿ three usage-based tiers, no per-seat option
⏺ WebFetch(northwind.com/changelog)
⎿ "Simpler, usage-based pricing" · 14 Jan 2026
⏺ WebSearch("northwind pricing change 2026")
⎿ 6 results · two teardowns, one HN thread
⏺ wire_search("northwind pricing")
⎿ 4 entries · $18/seat list, Mar 2025
⏺ wire_write(entry)
⎿ saved to competitors · tagged pricing, 2026-q1
Yes, and it is a bigger change than the changelog makes it sound. On 14 January they retired per-seat entirely and moved to three usage-based tiers billed on monthly active records, from $0.04 per thousand with volume breaks at 5M and 50M.
The headline effect is that our 20% advantage is gone at list. We were comparing $14.40 against $18 a seat, which held while both sides counted seats. The comparison now depends on how much an account processes, so there is no single number left to put on a slide.
I have filed it with the tier table and both teardowns linked.
❯ where does the crossover land?
⏺ wire_navigate("northwind-pricing-2026")
⎿ 3 related · our tiers, discount policy, Q4 win/loss
Around thirty seats for a typical account, though it tracks usage more than headcount. Below that we are still meaningfully cheaper and the gap widens as accounts get smaller, because their first tier has a floor we do not. Above it they pull ahead, and by a hundred seats they are about 15% under us unless the account is unusually write-heavy.
Worth flagging: the Q4 win/loss notes say price came up in four of nineteen competitive losses to them. It may not be the wedge we think.
wire · 5 tools connected research filed
Codex
Battle cards
Q3 roadmap
Churn analysis
wire · connected
Battle cards
Northwind isn't 20% pricier anymore. Rework their battle card.
wire_search "northwind pricing" 5
wire_navigate "northwind-pricing-2026" 3
I have read the pricing entry your Claude session filed on the 14th,
along with the tier table and the Q4 win/loss notes it links to.
Price is not going to carry this card any more. At list we are ahead
under about thirty seats and behind above it, which is too conditional
to hand a rep thirty seconds before a call. The win/loss data agrees:
price came up in four of nineteen competitive losses last quarter.
Reliability is the obvious replacement, but I do not want to claim it
without something recent and specific. Shall I go and find it?
yes, find something recent
WebSearch "northwind outage" 9
WebFetch status.northwind.com/history
wire_write uptime · 3 incidents this quarter
Update battlecards/northwind.md +18 −6
Found it, and it is better than I expected.
Their status page shows 7h 12m of full downtime last Wednesday, the
12th. That is their third incident this quarter and the second over
four hours. Rolling ninety-day availability comes out at 99.1% against
the 99.9% in their public SLA.
I have saved the incident history to Wire and rewritten the card
around it. Price now sits in a "when it comes up" section near the
bottom with the thirty-seat crossover spelled out.
Ask anything gpt-5.6
9:41
Sales bot connected to Wire
acme call in 10, they're evaluating northwind too
Don't lead on price. Northwind closed that gap on 14 January and it only holds under about thirty seats. Acme is at forty-five.
Lead on reliability instead. Northwind was down 7h 12m last Wednesday, third incident this quarter, and rolling availability is 99.1% against a 99.9% SLA.
exact date?
12 February. Second one over four hours this quarter.
perfect 🙏
Message
Code has GitHub. Everything else has Wire.
Anything goes in.
Agent-ready context comes out.
Upload files, let agents write entries directly, or push through the API. It all lands in one container that any MCP client can read.
Add context
Files .pdf .csv .docx .md
Agents wire_write
API REST API
Connectors 
├─ features.csv
├─ teardown.pptx
├─ api:weekly-scrape
└─ agent:analysis
├─ teardown.pptx
├─ api:weekly-scrape
└─ agent:analysis
Access anywhere
https://wire-public.mcp.usewire.io/container/Kp4oRo16CXk8YBfT/mcp
Works with 
+ any MCP client
+ any MCP client Containers
Context that holds together.
Each container is an isolated, permissioned environment for context, with its own database, MCP server, and API. Spin one up per person, team, project, or anything else. Every connected agent gets scoped access to only what it needs.
4,218 entries
├─ accounts.xlsx
├─ contracts.pdf
├─ agent:meeting-notes
└─ api:crm-sync
├─ contracts.pdf
├─ agent:meeting-notes
└─ api:crm-sync
932 entries
├─ features.csv
├─ teardown.pptx
├─ api:weekly-scrape
└─ agent:analysis
├─ teardown.pptx
├─ api:weekly-scrape
└─ agent:analysis
13,412 entries
├─ getting-started.md
├─ events.jsonl
└─ metrics.tsv
├─ events.jsonl
└─ metrics.tsv
Private by default
Every container is locked down by default. Connected agents only see what they have been scoped to.
Optionally public
Flip a container to public for read-only access when sharing openly makes sense. Any agent can connect.
Spin up, swap, delete
Create a container in seconds. Delete it when it is no longer needed. Connect an agent to whichever container fits the work.
Collaboration
Not one after another. All at once.
The handoff above happens one agent at a time, but a container does not make anyone wait their turn. Connect as many people and agents as the work calls for, each scoped to only what they should see, all reading the same context as it changes.
competitor-intel 932 entries
Intelligence
Graphs, not a file dump.
Wire analyzes everything inside a container and builds connections across it. Semantic neighbors, related claims, entity links, composites, time. Agents pivot from any result through whichever edges fit the question.
Privacy and security
Private by default. Provable on demand.
Four questions decide whether you can trust an agent with something that matters: what it can reach, who allowed it, whether you can prove it, and whether you can take it back. Every container answers all four.
What can it access?
Isolated
Every context container has its own storage and MCP endpoint. Containers do not share state. Credentials are scoped per container, and none of them escalate to other containers, other users, or org-wide access. The answer is the container boundary itself.
Who authorized it?
Permissioned
Each container has its own policies. Credentials are scoped per container, with a per-credential tool allowlist on top. Your agent only ever sees the tools you granted, on the container the credential scopes to. Access exists because someone granted it, and the grant is on record.
Can you prove it?
Auditable
Two layers are on the record: every action an agent takes through a scoped credential, and every management change (creation, sharing, archival, member changes). What was done, when, on whose authority. Compliance evidence, incident response, and customer-facing transparency come from the same trail.
Can you revoke it?
Revocable
Revoke a credential and access ends, without redeploying the agent or touching anything else. Sharing is explicit and per container, so anything that was granted can be ungranted, one agent at a time or all at once.
Need self-deployment, region pinning, or a custom DPA? Talk to us.
Frequently asked questions
Everything you need to know about Wire.
Still have questions? Get in touchWhat is Wire?
You decide something with one agent, switch to another, and it does not know. Wire is one container that every agent you work with reads and writes, so what one of them works out the next one already has. It holds documents, structured data, and entries agents write themselves, and any tool that speaks MCP can reach it.
How do I share context between different AI agents?
Point them at the same container. Each agent connects over MCP with its own scoped credential, reads what the others have written, and writes back into the same place. Nothing is exported or copied between tools, so there is one version of your context rather than a stale copy sitting in each one.
What happens when I switch from one agent to another?
Nothing, which is the point. The second agent reads the same container the first one wrote to, so decisions, findings and corrections carry over without being re-explained. That also works across people: a colleague connecting their own agent to a shared container picks up work they never saw happen.
How does it work?
Create a container for a project, team, or anything else an agent needs to know about. Add context through file uploads or let agents write entries directly. Wire organizes and structures everything into AI-optimized context. Connect from Claude, Cursor, or any MCP-compatible tool, or wire it into your own agent through the SDK.
What file types do you support?
Wire supports common document formats including PDF, Word documents, text files, markdown, CSV, JSON, and more. See the full list of supported file types.
What AI tools work with Wire?
Wire works with any AI tool that supports MCP (Model Context Protocol). This includes Claude Desktop, Claude Code, Cursor, Cline, and more. See setup guides for all supported tools.
What is a context container?
Each container is an isolated, permissioned environment for context, with its own database, MCP server, and API. Wire organizes and structures whatever lives inside, so connected agents get scoped access without manual setup. Containers adapt to their content.
Can I share containers with my team?
Yes. Invite people to an organization to share containers, each with scoped access. The same container can also connect to many agents at the same time, so every connection, human or agent, works from the same context.
Can I use Wire in an agent I'm building?
Yes. The Wire SDK gives your agent a drop-in connect flow. The user authorizes once and the SDK hands back a scoped MCP endpoint and key. From there your agent reads and writes the container through MCP or the REST API. Builders and agencies use this to skip building a context layer for every agent or client. See Build with Wire, or the Agent Products and AI Agencies solution pages.
Is my data private?
Yes. All containers are private by default and only accessible to you and the team members you explicitly invite. We never share your data with third parties or use it to train AI models.
How does pricing work?
Every new organization gets 1,000 starter credits to try Wire. Pick a plan for monthly credits and the option to buy additional credit packs. All plans include unlimited containers, team members, and MCP access. See our pricing page for details.
From the blog
Context engineering, AI agents, and what we're learning.
Pruning context made agents more accurate, not just cheaper
Pruning an agent's tool history raised task completion from 71% to 91.6%. What the summary rescued was not content but the agent's place in its own work.
What chunk size should you use for RAG?
Studies disagree on the best chunk size for RAG: one found 100 tokens with zero overlap won. What the research supports, and the bigger lever most guides miss.
For teams shipping agents
Building an agent? Don't build the part underneath.
Every user gets their own container. Ingest, scope, permissions, retrieval, sync and audit ship with it, reachable over MCP or REST.