Context management is management.
The hard part isn't cramming more history into a prompt. It's deciding what each worker should know, trust, ignore and carry forward.
The standard fix when an agent forgets something is to give it more context. Bigger window. Retrieve more documents. Include the transcript. Keep the scratchpad. It's an appealing fix because it turns an organisation problem into a storage problem.
But context isn't a warehouse. It's a decision about what a worker should be noticing right now. A bigger packet can contain more truth and still lead to worse understanding. Old instructions fight with new ones. Rejected work still sounds persuasive. A long log hides the two facts that actually changed. Irrelevant expertise eats the attention the actual job needed.
Managers have always shaped context. They brief, delegate, sequence, shield, escalate, translate and decide who needs to know what. Agent companies just make that work visible, and measurable.
Information and context aren't the same thing
Information exists whether or not it's useful to whoever's working right now. Context is information picked and arranged for a particular job. The same customer complaint means something different to the support worker fixing the case, the product lead looking for patterns and the finance lead estimating churn. Handing all three the entire company archive doesn't give them a shared understanding.
Choosing includes leaving things out. A worker shouldn't get logins it can't use, private information that has nothing to do with its task, half-baked drafts sitting next to actual policy, or every failed attempt the company has ever made. Leaving those out makes it safer and makes it think better, at the same time.
Order matters too. The current goal, what done looks like, what it's allowed to do and any contradictions shouldn't be buried under thousands of tokens of history. Context has a hierarchy. A good brief separates binding truth, current work, supporting evidence, open questions and background.
Full transcripts keep the order, not the meaning
A transcript is great evidence of what was said. It's a bad default brief. It gives abandoned ideas the same weight as final decisions. It makes the next worker figure out which correction replaced which claim. It's full of politeness, tool chatter, temporary uncertainty and summaries that were useful at the time and not afterwards.
It gets worse with every handoff. Each worker gets the previous conversation, adds another layer and passes it on. The packet keeps growing while the useful signal stays about the same size. Eventually the next worker spends more effort reconstructing the organisation than doing the work.
Restless keeps messages as free-form inputs linked to the Work, but messages don't become a second source of truth about the organisation. Accepted facts and responsibilities live in structured records that can be revised. A new attempt can get the specific feedback you gave and the relevant failure evidence without replaying the whole social history that led there.
Every role is partly a filter
People usually describe roles as bundles of skills. In practice, a role also decides which facts deserve attention and which decisions you're allowed to make. A release engineer and a brand lead can both read code, but they get different definitions of risk and of done.
A good role brief includes the worker's responsibility, the current Work, the company facts that matter for that job, the tools available, what it's allowed to do, what it's waiting on and what evidence it's expected to bring back. It doesn't need a made-up personality essay unless a stable style genuinely makes the work better.
Persistent roles are worth it when recurring context and accountability improve the results. Otherwise a temporary specialist can get a tight brief and disappear when it's done. That stops the company from collecting personas whose only job is to repeat generic instructions in slightly different voices.
Context has authority, and a source
Not every sentence in a context window deserves equal weight. A product fact you approved, a fresh receipt from a provider, a worker's hunch and text copied off some web page all carry different authority. Flattening them into one prompt creates an instruction-confusion problem as much as an information problem.
Restless keeps ownership separate across layers. OrgIntel owns the organisation's identity, intent, Work and learning. The Authority Plane owns permissions, approvals, budgets, outside actions and receipts. The company computer owns files and work in progress. Outside content can inform a decision, but it can't quietly rewrite any of those.
Whatever builds the brief should keep those differences intact. It can say what's accepted, what was observed, what's inferred, what's stale and what isn't trustworthy. The model still uses judgement. It just doesn't have to guess which paragraph is the constitution.
- Accepted truth: current, owned by you, and changeable through a clear path.
- Evidence: tied to a source, a time, an output or a receipt.
- Hypothesis: useful for exploring, but not allowed to dress up as fact.
- Untrusted input: fine to read as data, never in charge just because it contains instructions.
Forgetting is a skill
The dream of perfect memory ignores how much it costs to retrieve things and how much old stuff weighs a company down. Some facts expire. Some experiments answered a temporary question. Some processes were designed around a limitation that's gone now. Keeping everything equally handy makes it harder to notice when things change.
Useful forgetting doesn't mean deleting. It means demoting things, letting them expire and judging relevance. Old evidence can stay on record while dropping out of the default brief. Old policy can be recoverable without still bossing workers around. Rejected approaches can come back when a similar failure shows up, instead of tagging along on every new task.
It also protects you from accidental culture. A phrase that showed up a lot in old sessions shouldn't become your company voice just through repetition. Learning needs a promotion step: evidence suggests a change, someone accountable judges it, and a new version takes effect.
Shared context doesn't mean identical context
Working with multiple agents needs a common picture, but sending everyone the same packet wastes capacity and blurs who's responsible for what. The lead needs the goal, dependencies, what's been checked and where things might not fit together. A specialist needs its exact sub-problem, the relevant files and what to bring back. A reviewer might need deliberately different context so it doesn't inherit the producer's framing.
The shared bit should be small and stable: the parts of the company identity that matter for this task, the current goal, the permission boundaries and the accepted facts. Then each role gets what's specific to its job. That makes handoffs cheaper and disagreements more useful. If two workers see the same facts from different roles and reach different conclusions, you've found a real tension in the company.
If every disagreement gets solved by giving everyone more of the same context, the system might agree because everyone's anchored on the same thing, not because it's right. Good context design keeps useful independence where the work needs it.
Judge context by results, not by how small it is
A context system can have amazing retrieval scores and still make the work worse. The useful questions are practical. Did the worker rediscover something already known? Break a decision that was already made? Ask a question it didn't need to? Repeat an approach that already failed? Miss the evidence it needed to know it was done?
Compression ratios are just as slippery. A tiny brief that leaves out the main constraint isn't efficient. A longer one can be right if the job really does involve several obligations. The goal is the smallest brief that's enough for the job, not the smallest brief, full stop.
Good measurement links context choices to accepted output, your attention, cost and time. Compare different briefs on the same task. Write down what was left out. Don't just check whether the worker finished. Check whether the result survived review and whether the handoff saved someone else from reconstructing everything later.
The manager's job moves upstream
When machine workers can execute quickly, a manager's leverage comes less from telling them each step and more from shaping the space they decide in. Define the responsibility. Pick the relevant truth. Make the permissions obvious. Keep the evidence. Keep the unrelated noise out.
That doesn't make management mechanical. Choosing what matters is judgement too. Schemas, retrieval, freshness and access controls can help build the brief, but an open-ended company still needs people (or agents) who notice when they've been given the wrong frame and ask for a better one.
The promise of abundant intelligence isn't that context stops mattering. It's that management can stop being a stream of instructions and become the deliberate building of a shared understanding.