Context != Context: One window, many lifecycles (AI Native DevCon, Nov 2026)

https://youtu.be/NxypyNu1944

Workflow instructions, coding conventions, tickets, wiki pages, and observability data all end up in the same context window, and we call all of it "context." To the model, it's tokens shaping the next prediction. For the people who write, maintain, and rely on context, these tokens differ a lot in where they come from, how much authority they carry, and how they can go wrong. Some are deliberately authored and quality assured. Others emerge as a byproduct of work and are dynamically retrieved and reconciled from organizational systems.

This talk discusses how these differences change the ways that each kind is created, checked, and delivered to the agent, and that treating them alike can lead to using unsuitable tools and processes. You will leave with a clearer idea of the types of context you have available for your coding agents, and the techniques to maintain and distribute them.

Transcript

1. GETUNBLOCKED.COM Context != Context One window, many life cycles
2. What is “context”? Context window: 36% left (171K used / 258K)
3. Context window AGENTS.md “Always use … Never do …” Task / intent / specification SKILL.md Spec workflow Feature discussion Wiki Past decision history Module architecture gh Slack Repo search docs/architecture.md
4. Context window AGENTS.md “Always use … Never do …” Task / intent / specification SKILL.md Spec workflow Feature discussion Wiki Past decision history Module architecture gh Slack Repo search docs/architecture.md
5. Codified AGENTS.md SKILL.md Ambient docs/arch.md Wiki Slack gh DELIVERY MECHANISM DELIVERY MECHANISM Deliberately authored Assembled at task time Versioned (and tested) Connects agent to the environment FOR CONTEXT WITH… FOR CONTEXT WITH… High reusability Low reusability Lower frequency of change High frequency of change
6. They complement each other: Example of a code review Codified AGENTS.md “Use exponential backoff”: follows it AGENTS.md “Log every retry”: missing log line Ambient Slack thread Billing retries were capped at 3 after last month’s incident. This sets 5. Ticket Scope is webhooks only. Payments client is out of scope.
7. Importance depends on the task Planning and design Investigating an incident Onboarding Implementation Code review Code in the repo Codified Ambient
8. Codified The “Context delivery life cycle” 1 Generate 3 Distribute Making implicit knowledge explicit Context as a package 1 3 2 4 2 Evaluate 4 Observe TDD for context Learn from use in the wild Generate Evaluate Distribute Observe How to write skills Skill evals Skill registries, plugins, descriptions Analyse agent traces, learn from errors
9. How does this work when the context is coming from the constantly changing environment? Ambient Generate Evaluate Distribute Observe ? ? ? ?
10. The goal: High quality context C Clarity A Actionability F Fidelity (S) Security E Efficiency
11. Ways to deliver ambient context to the agent MCP server Agent
12. What’s responsible for efficiency, fidelity, security? Atlassian MCP GitHub MCP Agent Datadog MCP Efficiency Slack MCP not called Docs MCP not called Security Who may see this? What’s relevant, what’s noise? Fidelity What’s stale, conflicting, unsupported?
13. Ways to deliver ambient context to the agent MCP server Context engine Agent sync retrieve Efficiency Reason, rank, reconcile Fidelity Security
14. The “Context delivery life cycle” for a context engine – Generate Ambient – Distribute Ingest Evaluate Assemble Observe Sync Evaluate the pipeline, not the context Retrieve, reason, rank, reconcile User feedback Link Index Cache Respect permissions Learn from agent traces
15. Ingest Evaluate Assemble Observe
16. INGEST Freshness and sync reliability are key
17. INGEST Some lessons from syncing data Variety of syncing mechanisms and data models Keeping up with changes efficiently Identity reconciliation Webhooks, polling, scheduled backfills Incident: 737k rewrites a day from PR changes Who is who across all the sources Different permission models
18. Ingest Evaluate Assemble Observe
19. EVALUATE Testing a skill vs testing an engine A skill eval EnterpriseRAG-Bench data set UNDER TEST The instructions the author wrote Retrieval and reasoning over a whole company TEST DATA A handful of test tasks to apply the skill to 500k documents, 9 data sources HARD CASES The edge cases the author thought of Conflicts, near-duplicates, answers that aren’t there GROUND TRUTH The author knows what good looks like Gold documents, gold answer, list of facts
20. Ingest Evaluate Assemble Observe
21. ASSEMBLE Tuning the engine loop Depth vs Latency Cost vs Quality Caller “personas” Coding agent MCP Code Review Chat
22. ASSEMBLE Conflict detection 13,742 source checks in one org across ~2 weeks 10% conflicts confirmed 92% conflicts settled 809 574 settled by checking the code settled another way 105 8 told the agent No verdict
23. Ingest Evaluate Assemble Observe
24. OBSERVE Conflict detection patterns: Improve the ambient context itself MOST OVERRULED DOCUMENT SOURCES IN ONE ORG Wiki (Confluence) 81% Tickets (Jira) 75% Chat (Slack) 72% Monitoring (Datadog) 59% Web 58% GitHub 50% One FAQ page on the wiki was overruled 15×
25. OBSERVE Cluster and prioritise: Learn from human feedback 875 downvoted answers May 2026 ~34% IGNORED WHO IS ASKING ~31% WRONG CONFLICT RESOLUTION ~31% MISSING CONTEXT A staff engineer and a new hire get the same answer E.g. a year-old Slack thread beating last week’s PR Some data never got ingested
26. Ways to deliver ambient context to the agent MCP server Context engine sync retrieve Reason, rank, reconcile Memory periodically Reason, rank, reconcile Agent
27. Maintaining the context itself
28. Think about where you put what Codified Don’t bother codifying frequently changing environmental data Ambient Don’t trust that key reusable instructions will emerge from ambient context
29. What’s a good “ambiance”? Context engines can handle Knowledge spread across many systems No common structure or categorization Outdated, conflicting information - as long as there are signals that help resolve (code, experts, …) Make sure knowledge is accessible to the engine Monitor for docs that impede conflict resolution We have to handle Write things down, think out loud
30. Sources of ambient context in the future - ? Today: Humans, communicating Agents PRs, tickets, chat, docs Systems Trending towards: Humans less of it, more value per item Human corrections e.g. in agent traces Agent-written content System “exhaust” PR descriptions, plans, memory telemetry, incidents
32. Thank you Birgitta Boeckeler · getunblocked.com